Comparison between chemometrics and machine learning for the prediction of macronutrients in cheese using Imaging spectroscopy

Traditional methods for assessing cheese's nutritional content are often labor-intensive, destructive, and environmentally taxing. This study explores the non-destructive spectral imaging technique, also called Hyperspectral Imaging (HSI) combined with chemometrics and machine learning (ML) to predict fat and protein content in 73 cheese samples. By adopting a broad-based approach, we integrated a diverse range of cheese varieties into a single model, aiming to enhance predictive accuracy. We evaluated multiple pretreatment methods, feature selection approaches, and model types to determine their predictive performance. Chemometric approaches, including Partial Least Squares (PLS) and its variants, were compared with ML models such as a Multilayer Perceptron (MLP), alongside variable selection techniques like CovSel and IPW-PLS. For protein prediction, the best-performing chemometric model used Extended Multiplicative Scatter Correction (EMSC) (degree = 6) pretreatment, achieving an R2_pred of 0.96, Mean Squared Error of Prediction (MSEP) of 2.61, and Standard Error of Prediction (SEP) of 1.61 without variable selection. The Uninformative Variable Elimination in PLS (UVE-PLS) model, also in the chemometric category, further enhanced accuracy with an R2_pred of 0.98, but required 80 selected variables. For fat prediction, the Iterative Predictor Weighting PLS (IPW-PLS) chemometric model with 15 selected variables achieved the highest R2_pred (0.94) and RMSEP (2.15). ML models, such as the MLP, performed comparably, with the best MLP model yielding an R2_pred of 0.94 for protein and 0.97 for fat, but without the benefit of interpretable variable selection. This highlights the advantage of chemometrics in providing practical insights into important wavelengths for fat and protein prediction.