Predicting the flavor potential of green coffee beans with machine learning-assisted visible/near-infrared hyperspectral imaging (Vis-NIR HSI): Batch effect removal and few-shot learning framework

Abstract

The current study investigates the potential of machine learning-assisted visible/near-infrared hyperspectral imaging (Vis-NIR HSI) for rapid and non-invasive evaluation of the overall flavor quality of green coffee beans. Spectral data was subjected to preprocessing and machine learning (ML) batch effect removal. Multivariate statistical analysis and various ML classifiers were compared and optimized for pass/fail flavor quality grading following non-few-shot and few-shot learning frameworks. Involving the full training dataset, the non-few-shot learning binary grading model achieved 100% accuracy using linear discrimination analysis (LDA). The few-shot learning model trained with data from only four of the 13 coffee bean types yielded the highest discrimination accuracies of 99% and 97%, in the internal validation and external evaluation, respectively, using the ensemble learning algorithm of LightGBM. Furthermore, the incorporation of ML batch effect removal enhanced the few-shot learning LightGBM model accuracy by 27% in external evaluation. The results demonstrate that assisted by ML batch effect removal and classification, Vis-NIR HSI serves as a rapid, sensitive, robust, and practical tool for evaluating the flavor potential of green coffee beans, especially when labeled data for training is limited.