关键词: 茶树, 儿茶素, 高光谱成像, 多任务学习, 连续投影算法, 叶片成熟度, 植物生理先验

Abstract:

【Objective】To meet the demand for in situ, real-time, and non-destructive monitoring of tea leaf quality in smart tea gardens, this study proposed a multi-component quantitative prediction method based on visible-near infrared hyperspectral imaging (400-1 000 nm) integrated with plant physiological priors, so as to realize rapid assessment of secondary metabolites, such as catechins.【Method】Twelve tea cultivars were used as experiment materials. Leaves were classed as young (2nd-3rd), mature (4th-5th), and old (6th-7th) to represent physiological gradients, giving 429 independent samples in total. For each leaf, hyperspectral reflectance was recorded, and the contents of catechin (C), epicatechin (EC), epigallocatechin (EGC), epicatechin gallate (ECG), epigallocatechin gallate (EGCG), gallocatechin (GC), gallocatechin gallate (GCG), caffeine (CAF), and gallic acid (GA) were quantified by high-performance liquid chromatography. Spectra were preprocessed using Savitzky-Golay (SG) smoothing, standard normal variate (SNV), and first derivative. The generalization performance of partial least squares regression (PLSR), support vector regression (SVR), XGBoost, multilayer perceptron (MLP), and 1D-CNN-SE was compared. Based on the best single-task model, a multi-task XGBoost framework was then developed to simultaneously predict the nine components. Additionally, the successive projections algorithm (SPA) was used to select key wavelengths from the full spectrum (557 bands) to assess the feasibility of simplified modeling.【Result】When leaf maturity was ignored, EGCG and GCG showed a spurious negative correlation due to confounding by maturity, which was a clear instance of Simpson's paradox. This meant that maturity needed to be accounted for, either by stratifying the data or by explicit adjustment, so that models would not learn false correlations. Among the tested models, XGBoost combined with SG+SNV+first-derivative preprocessing achieved the best performance on the small-sample with high-dimensional spectral dataset. The R² values for EGCG and GCG were approximately 0.60 and 0.59, respectively. In contrast, 1D-CNN-SE achieved calibration set R² value above 0.90 but dropped to about 0.40-0.50 on the prediction set, showing clear overfitting. The multi-task XGBoost model predicted all nine components in a single run; the R² value for EGCG remained about 0.57, and the RPD values for EGCG, GCG, and CAF were about 1.48-1.52, with only a slight loss in single-component accuracy but a substantial gain in throughput. After SPA reduced the full spectrum from 557 bands to 18 key wavelengths, the R² for EGCG remained about 0.54. These key wavelengths fell mainly in the visible region and the near-infrared overtone absorption bands, supporting the design of portable quality sensors based on discrete wavelengths.【Conclusion】The spectral modeling that integrated physiological priors, such as leaf maturity, could enhance the model’s explanatory power and improve its generalization across cultivars and leaf ages. The single- and multi-task XGBoost frameworks, combined with wavelength selection, provided a theoretical basis and practical support for rapid multi-index tea quality assessment in tea plantations.

Key words: tea plant, catechins, hyperspectral imaging, multi-task learning, successive projections algorithm (SPA), leaf maturity, plant physiological priors