Band mapping in hyperspectral images for predicting physicochemical parameters in black tea (Camellia sinensis)

Rodríguez-León A.; Oblitas J. · AIMS Agriculture and Food (2026) · DOI: 10.3934/agrfood.2026029

Resumen

Black tea is one of the most widely consumed infusions worldwide, whose quality changes during storage due to physicochemical and color changes. However, quantitative prediction of key physicochemical parameters, such as pH and water activity (Aw), during storage has been scarcely explored using hyperspectral imaging and machine learning. This study evaluated visible–near infrared (VIS–NIR) spectra acquired by hyperspectral imaging (HSI) from samples stored under different conditions to predict pH and Aw. A reproducible machine learning pipeline (second-order Savitzky–Golay derivative + standard normal variant (SNV), train–test baseline evaluation followed by stratified cross-validation) was implemented, and PLS, elastic net, random forest, support vector regression (SVR)-radial basis function (RBF), and XGBoost were compared. SVR-RBF showed the best performance (pH R2 ≈ 0.94, Aw R2 ≈ 0.82). Consensus-based spectral attribution identified relevant bands in the visible (390–550 nm), associated with polyphenolic chromophores, and in the short NIR (870–960 and 980–1000 nm), linked to O–H and C–H overtones. Leave-one-band-out (LOBO) validation revealed minimal variations (ΔR2 ≈ 0), confirming informational redundancy and robustness against band exclusion. In conclusion, VIS–NIR combined with machine learning is a potentially effective and portable tool for monitoring critical quality parameters in black tea, reducing the need for recalibration and facilitating the implementation of compact sensors in at-line and on-line applications.

Instituciones: Universidad Nacional de Cajamarca; Universidad Privada del Norte