Rapid discrimination of different primary processing Arabica coffee beans using FT-IR and machine learning

Abstract

In this study, fourier transform infrared spectroscopy (FT-IR) analysis was combined with machine learning, while various analytical techniques such as colorimetry, low-field nuclear magnetic resonance spectroscopy, scanning electron microscope, two-dimensional correlation spectroscopy (2D-COS), and multivariate statistical analysis were employed to rapidly distinguish and compare three different primary processed Arabica coffee beans. The results showed that the sun-exposed processed beans (SPB) exhibited the highest total color difference value and the largest pore size. Meanwhile, the wet-processed beans (WPB) retained the most bound and immobilized water in green and roast coffee beans. The FT-IR data analysis results indicated that the functional group composition was similar across the three different primary processed coffee beans, while significant differences in structural characteristics were observed in 2D-COS. The multivariate statistical analysis demonstrated that the orthogonal partial least squares-discriminant analysis model could effectively distinguish the different types of coffee beans. The machine learning results indicated that the six models could rapidly identify different samples of primary processed coffee beans. Notably, the SNV-Voting model demonstrated superior predictive performance, with an average precision, recall, and F1-score of 88.67%, 88.67%, and 0.88 for three primary processing coffee beans, respectively.