Classification and Authentication of Edible Oil-Based Products: A Comparison of Machine Learning Algorithms Combined with FTIR Spectroscopy

This study assessed the value of FTIR spectroscopy combined with machine learning (ML) algorithms in the identification and categorization of pure edible oils, including njangsa seed oil (NSO), palm kernel oil (PKO), coconut oil (CCO), njangsa seed-palm kernel oil (NSOPKO), and their formulated margarines (NSOPKO and NSOCCO). The magnitude of adulteration in each oil and margarine was also quantified using ML regression models, with sunflower oil and canola-flax seed oil margarine as adulterants. The fingerprints of the oils and margarines, extracted in the spectral region of 4000 - 600 cm⁻¹, were integrated with ML models. The first two principal components elucidated 99.4% and 98% of the variance in pure oils and margarines, and 90.1%, 88.3%, 88%, 97.3%, and 98.3% of adulterated PKO, NSO, CCO, NSOCCO, and NSOPKO, respectively, facilitating visualization. Pure margarines were accurately classified (100%) in all models. The most effective models in classifying pure oil were KNN (97%), followed by LR (93%), SVM (83%), LightGBM (53%), and DT (50%). The R² values obtained from all models for adulterated PKO, NSO, CCO, NSOPKO, and NSOCCO ranged from 59% to 99%, 55% to 99%, 45% to 94%, 69% to 98%, and 59% to 94%, respectively. SVM and DT exhibited lower performance, while KNN emerged as the best model. The FTIR spectroscopic technique eliminated the need for laborious sample preparation. The ML methods, in conjunction with FTIR spectroscopy, can effectively discriminate and quantify adulterants in oil and margarine. This approach holds promise for improvement in quality control settings, facilitating rapid authentication of similar products and broad adoption.