Malavi Derick, Raes Katleen, Van Haute Sam
Department of Food Technology, Safety and Health, Faculty of Bioscience Engineering, Ghent University, Coupure Links 653, 9000, Ghent, Belgium.
Center for Food Biotechnology and Microbiology, Ghent University Global Campus, 119, Songdomunhwa-Ro, Yeonsu-Gu, Incheon, 21985, South Korea.
Curr Res Food Sci. 2024 Oct 29;9:100913. doi: 10.1016/j.crfs.2024.100913. eCollection 2024.
Detecting adulteration in extra virgin olive oil (EVOO) is particularly challenging with oils of similar chemical composition. This study applies near-infrared hyperspectral imaging (NIR-HSI) and machine learning (ML) to detect EVOO adulteration with hazelnut, refined olive, and olive pomace oils at various concentrations (1%, 5%, 10%, 20%, 40%, and 100% m/m). Savitzky-Golay filtering, first and second derivatives, multiplicative scatter correction (MSC), standard normal variate (SNV), and their combinations were used to preprocess the spectral data, with Principal Component Analysis (PCA) reducing dimensionality. Classification was performed using Partial Least Squares-Discriminant Analysis (PLS-DA) and ML algorithms, including k-Nearest Neighbors (k-NN), Naïve Bayes, Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Networks (ANN). PLS-DA, k-NN, RF, SVM, NB, and ANN models achieved accuracy rates of 97.0-99.0%, 96.2-100%, 96.5-100%, 98.6-99.5%, 93.9-99.7%, and 99.2-100%, respectively, in discriminating between pure EVOO, adulterants, and adulterated oils. PLS-DA, RF, SVM, and ANN significantly outperformed Naïve Bayes (p < 0.05) in binary classification, with Matthews correlation coefficient (MCC) values exceeding 0.90. All the binary classifiers except Naïve Bayes, when coupled with SNV/MSC, Savitzky-Golay smoothing and derivatives, consistently achieved perfect scores (1.0) for accuracy, sensitivity, specificity, F1 score, precision, and MCC in distinguishing pure EVOO from adulterated oils. No significant differences (p > 0.05) in model performance were found between those using full spectra and those based on key variable selection. However, PLS-DA and ANN significantly outperformed k-NN, RF, and SVM (p < 0.05), with MCC values ranging from 0.95 to 1.00, indicating superior classification performance. These findings demonstrate that combining NIR-HSI with machine learning, along with key variable selection, potentially offers an effective, non-destructive solution for detecting adulteration in EVOO and combating fraud in the olive oil industry.
对于化学成分相似的油类，检测特级初榨橄榄油（EVOO）中的掺假情况极具挑战性。本研究应用近红外高光谱成像（NIR-HSI）和机器学习（ML）来检测EVOO与榛子油、精炼橄榄油和橄榄果渣油在不同浓度（1%、5%、10%、20%、40%和100% m/m）下的掺假情况。使用Savitzky-Golay滤波、一阶和二阶导数、乘法散射校正（MSC）、标准正态变量变换（SNV）及其组合对光谱数据进行预处理，并通过主成分分析（PCA）进行降维。使用偏最小二乘判别分析（PLS-DA）和机器学习算法（包括k近邻算法（k-NN）、朴素贝叶斯算法、随机森林（RF）、支持向量机（SVM）和人工神经网络（ANN））进行分类。在区分纯EVOO、掺假物和掺假油方面，PLS-DA、k-NN、RF、SVM、NB和ANN模型的准确率分别达到97.0 - 99.0%、96.2 - 100%、96.5 - 100%、98.6 - 99.5%、93.9 - 99.7%和99.2 - 100%。在二分类中，PLS-DA、RF、SVM和ANN的表现显著优于朴素贝叶斯算法（p < 0.05），马修斯相关系数（MCC）值超过0.90。除朴素贝叶斯算法外，所有二分类器在与SNV/MSC、Savitzky-Golay平滑和导数相结合时，在区分纯EVOO和掺假油时的准确率、灵敏度、特异性、F1分数、精度和MCC始终达到满分（1.0）。在使用全光谱的模型和基于关键变量选择的模型之间，未发现模型性能存在显著差异（p > 0.05）。然而，PLS-DA和ANN的表现显著优于k-NN、RF和SVM（p < 0.05），MCC值在0.95至1.00之间，表明其具有卓越的分类性能。这些发现表明，将NIR-HSI与机器学习以及关键变量选择相结合，有可能为检测EVOO中的掺假情况和打击橄榄油行业的欺诈行为提供一种有效的非破坏性解决方案。
