Scholarlink

Food fraud negatively impacts food systems and consumer trust. Among foods, extra virgin olive oil (EVOO) is particularly susceptible to adulteration. Among the various detection methods, spectroscopic techniques have seen considerable commercial success. However, their performances are heavily affected by human decisions during development, e.g. feature engineering. This research focuses on the use of near infrared (NIR) spectroscopy in combination with deep learning strategies to identify and quantify EVOO adulteration. The study employs two analytical approaches, namely: Partial Least Square (PLS) and Convolutional Neural Network (CNN) algorithms, which represent classical and deep chemometrics, respectively. Adulterated samples were prepared by blending EVOO with four seed oils (peanut, sunflower, maize, and soy) at concentration of 0, 0.5, 1.5, 3, 5, 10, 15, 20, and 100% (w/w). Spectra were acquired using both Fourier transform near-infrared spectroscopy (FT-NIR; 1000–2500 nm) visible spectroscopy (Vis; 380–900 nm). PLS models performed well for FT-NIR (R2 ~= 0.99, BIAS ≤ 0.17%, and RMSE ranging from 0.74 to 1.93%); However, the Vis models exhibited less satisfactory performances, with an R2 ≥ 0.93, BIAS ≤ 0.5%, and RMSE ranging from 2.00 to 3.68%. CNN models exceptional performances with both FT-NIR (BIAS ≤ 0.42%; and RMSE ranging from 1.09 to 1.51%) and Vis (BIAS ranging from 0.04 to 0.31%; RMSE close to 1.00%; and R2 > 0.99).

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Learning Versus Chemometrics: A Comparative Study in EVOO Adulteration Detection

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Andrea Bandiera,
  • Valentina Rosati,
  • Domenico Capone,
  • Alessandro Benelli,
  • Eleonora Taormina,
  • Riccardo Massantini,
  • Roberto Moscetti

摘要

Food fraud negatively impacts food systems and consumer trust. Among foods, extra virgin olive oil (EVOO) is particularly susceptible to adulteration. Among the various detection methods, spectroscopic techniques have seen considerable commercial success. However, their performances are heavily affected by human decisions during development, e.g. feature engineering. This research focuses on the use of near infrared (NIR) spectroscopy in combination with deep learning strategies to identify and quantify EVOO adulteration. The study employs two analytical approaches, namely: Partial Least Square (PLS) and Convolutional Neural Network (CNN) algorithms, which represent classical and deep chemometrics, respectively. Adulterated samples were prepared by blending EVOO with four seed oils (peanut, sunflower, maize, and soy) at concentration of 0, 0.5, 1.5, 3, 5, 10, 15, 20, and 100% (w/w). Spectra were acquired using both Fourier transform near-infrared spectroscopy (FT-NIR; 1000–2500 nm) visible spectroscopy (Vis; 380–900 nm). PLS models performed well for FT-NIR (R2 ~= 0.99, BIAS ≤ 0.17%, and RMSE ranging from 0.74 to 1.93%); However, the Vis models exhibited less satisfactory performances, with an R2 ≥ 0.93, BIAS ≤ 0.5%, and RMSE ranging from 2.00 to 3.68%. CNN models exceptional performances with both FT-NIR (BIAS ≤ 0.42%; and RMSE ranging from 1.09 to 1.51%) and Vis (BIAS ranging from 0.04 to 0.31%; RMSE close to 1.00%; and R2 > 0.99).