Fig. 3: Results of variable selection algorithms and machine learning models for peroxide value (IV) prediction in crude palm oil using Raman spectroscopy.

CARS variable selection results: A variable selection process showing RMSECV values and the number of retained variables, B distribution of selected variables in the full Raman spectrum, C 3D response surface for SVM hyperparameter optimization, D scatter plot of CARS-SVM model predictions versus reference values, E scatter plot of CARS-RF model predictions versus reference values. UVE variable selection results: F reliability plot distinguishing relevant variables (yellow) from random variables (red), G distribution of selected variables, H 3D response surface for SVM hyperparameter optimization, I UVE-SVM prediction scatter plot, J UVE-RF prediction scatter plot. GA variable selection results: K frequency distribution of selected variables, L distribution of selected variables in the spectrum, M 3D response surface for SVM hyperparameter optimization, N GA-SVM prediction scatter plot, O GA-RF prediction scatter plot. Across all three variable selection strategies, the majority of selected variables are concentrated in chemically meaningful Raman regions associated with lipid unsaturation and oxidation, particularly 1287–1657 cm⁻¹ (C–H bending and C=C stretching) and the carbonyl-related region around ~1748 cm⁻¹, which underpin peroxide value prediction.