Machine learning-assisted Raman spectroscopy for non-destructive analysis of crude palm oil quality

Area of Science:

• Analytical Chemistry
• Chemometrics
• Spectroscopy

Background:

• Crude palm oil quality assessment is vital globally.
• Traditional methods are often slow, destructive, and unsuitable for resource-poor settings.
• Developing rapid, non-destructive techniques is crucial for the palm oil industry.

Purpose of the Study:

• To develop and validate a machine learning-assisted Raman spectroscopy method for non-destructive quality assessment of crude palm oil.
• To predict key quality parameters: peroxide value (PV) and iodine value (IV).
• To identify spectral features indicative of oil quality.

Main Methods:

• Collected Raman spectra from 200 crude palm oil samples from Ghana.
• Applied second derivative preprocessing to enhance spectral features.
• Developed 12 predictive models using variable selection (CARS, GA, UVE) and regression methods (PLS, SVM, RF).
• Evaluated model performance using correlation coefficient (Rp) and ratio of performance to deviation (RPD).

Main Results:

• The Genetic Algorithm-Random Forest (GA-RF) model achieved high prediction accuracy for PV (Rp = 0.9831) and IV (Rp = 0.9752).
• The GA-RF model showed excellent predictive ability with RPD values of 7.7397 for PV and 6.3927 for IV.
• Identified key spectral regions linked to unsaturation (1287-1657 cm⁻¹) and oxidation (1748-1840 cm⁻¹).

Conclusions:

• Machine learning-assisted Raman spectroscopy provides a rapid and non-destructive method for crude palm oil quality assessment.
• This approach is suitable for resource-poor areas and has broad applications in the palm oil value chain.
• The GA-RF model demonstrates significant potential for real-time quality monitoring.