Comparing deep and classical Chemometrics: can CNN enhance the accuracy of EVOO adulteration detection from spectral data?

Abstract

Extra virgin olive oil (EVOO) has a high economic value and is therefore susceptible to adulteration with oils of lower quality and price. Spectroscopy, although not recognized as an official analysis methodology in EVOO, can be used in rapid screening to detect adulterants. Predictive models are usually developed with classical chemometrics, which require human knowledge in feature engineering. Deep chemometrics overcome human intervention by relying on neural networks. This study compares the use of PLS (Partial Least Squares) and CNN (Convolutional Neural Network) algorithm in combination with FT-NIR (1000–2500 nm) and Vis-NIR (380–900 nm) spectroscopy to predict EVOO adulteration using four different seed oils (peanut, maize, sunflower and soya). Adulterant concentrations at 0.5 % and 1.5 % were difficult to distinguish, as the subtle spectral changes were often masked by a low signal-to-noise ratio mainly due to high spectral similarity with the pure EVOO, instrumental noise, and intrinsic oil variability. The FT-NIR-based regressions generally showed minimal performance differences between PLS and CNN, regardless of the application of spectral pretreatment or data augmentation (RMSEP = 0.99–2.08 %), indicating that for this spectral range, the added complexity of CNN offered no significant advantage. Only the model obtained using CNN and FT-NIR for peanut adulteration was not able to converge. In contrast, the Vis-NIR models based on CNN significantly outperformed the PLS models, regardless of the use of pretreatment or data augmentation. Therefore, in the present study, deep chemometrics proved not to be a universal replacement for classical chemometrics, but rather a complementary tool that demonstrates its true value where the classical approach is less effective.