Measurement of four main catechins content in green tea based on visible and near-infrared spectroscopy using optimized machine learning algorithm

Abstract

The catechin content is critical for assessing the nutritional value and commercial worth of green tea, making detection methods valuable for its quality evaluation. This study utilized visible and near-infrared spectroscopy (Vis-NIRS, 400–2500 nm) combined with an optimized machine learning model, integrating the global search whale optimization algorithm (GSWOA) and kernel extreme learning machine (KELM), to predict the content of four main catechins (EC, ECG, EGC, and EGCG) in fresh tea leaves. Multiple preprocessing and effective wavelength selection methods were applied to improve prediction accuracy. The results showed that the distribution of effective wavelengths verified the importance of the visible spectrum. The GSWOA-KELM demonstrated superior performance, optimizing the modeling capabilities of KELM and outperforming other models, including partial least squares regression (PLSR) and random forest (RF), with $R_P^2$ values exceeding 0.98 and RMSEP values below 3.3 for all catechins. Notably, even without additional data processing, the model maintained good effects, achieving a minimum $R_p^2$ of 0.9751 and a maximum RMSEP of 3.3654. Therefore, the proposed approach proved to be both accurate and rapid for non-destructive, online catechin analysis and had practical value for tea appreciation.