Catechin is a major reactive substance involved in black tea fermentation. It has a determinant effect on the final quality and taste of made teas. In this study, we applied hyperspectral technology with the chemometrics method and used different pretreatment and variable filtering algorithms to reduce noise interference. After reduction of the spectral data dimensions by principal component analysis (PCA), an optimal prediction model for catechin content was constructed, followed by visual analysis of catechin content when fermenting leaves for different periods of time. The results showed that zero mean normalization (Z-score), multiplicative scatter correction (MSC), and standard normal variate (SNV) can effectively improve model accuracy; while the shuffled frog leaping algorithm (SFLA), the variable combination population analysis genetic algorithm (VCPA-GA), and variable combination population analysis iteratively retaining informative variables (VCPA-IRIV) can significantly reduce spectral data and enhance the calculation speed of the model. We found that nonlinear models performed better than linear ones. The prediction accuracy for the total amount of catechins and for epicatechin gallate (ECG) of the extreme learning machine (ELM), based on optimal variables, reached 0.989 and 0.994, respectively, and the prediction accuracy for EGC, C, EC, and EGCG of the content support vector regression (SVR) models reached 0.972, 0.993, 0.990, and 0.994, respectively. The optimal model offers accurate prediction, and visual analysis can determine the distribution of the catechin content when fermenting leaves for different fermentation periods. The findings provide significant reference material for intelligent digital assessment of black tea during processing.

Amino acids (AAs) and tea polyphenols (TPs) are essential quality indicators in tea, impacting sensory attributes and economic value. Hyperspectral technology enables efficient, real-time detection of these compounds on field-grown tea leaves. “The original spectra were preprocessed using fractional-order derivatives (0.1–1.0 orders) to enhance subtle spectral features. Compared to fixed integer-order derivatives (e.g., first or second order), fractional-order derivatives allow continuous tuning between 0 and 1, thereby amplifying minor absorption peaks while effectively suppressing noise amplification”. The Competitive Adaptive Reweighted Sampling (CARS) method selects optimal spectral bands, and Partial Least Squares Regression (PLSR) models were built with raw spectral reflectance as independent variables and AA and TP content as dependent variables. Results show that FOD had better prediction accuracy compared to classical integer-order derivatives, e.g., the optimal FOD order of 0.7 for AA prediction increased the R² from 0.73 to 0.80 and reduced the RMSE from 0.30% to 0.25%, while for TP prediction, a FOD order of 0.1 raised the R² from 0.40 to 0.42 and lowered the RMSE from 4.03% to 3.96%. In addition, CARS shows a better performance over the correlation coefficient (CC) method in model accuracy, contributing to more accurate selection of sensitive bands for the content prediction of tea ingredients. Our FOD–CARS–PLSR models achieved an R² of 0.80 and RMSE of 0.25% for AAs, and an R² of 0.42 and RMSE of 3.96% for TPs in fresh tea leaves. Beyond tea quality monitoring, this flexible preprocessing and modeling framework can be readily adapted to estimate biochemical or biophysical properties in other crops, soils, or vegetated ecosystems, offering a generalizable tool for precision agriculture and environmental sensing.