FB-CARS–PER-BHNet for Nondestructive Hyperspectral Assessment of Quality Indicators in Finished Black Tea

Nondestructive assessment of finished black tea is challenged by limited reference samples and highly correlated spectra. This study examined whether compact wavelength selection and selective residual modeling could support the prediction of chemically distinct quality constituents. We developed FB-CARS–PER-BHNet, combining frequency-aggregated, block-constrained competitive adaptive reweighted sampling (FB-CARS), a partial least squares regression (PLSR) backbone, and evidence-gated residual learning. Forty-seven finished black tea samples were represented by 243 visible–near-infrared wavelengths (401.05–999.20 nm) and evaluated by outer leave-one-out validation for water extract, epigallocatechin gallate (EGCG), epicatechin gallate (ECG), and lysine (Lys). Held-out R2 was 0.9068 for water extract, 0.9051 for EGCG, 0.8644 for ECG, and 0.7553 for Lys. FB-CARS reduced spectral redundancy, while compact PLSR models approached full-spectrum accuracy using 20, 25, 15, and five wavelengths, removing 89.7–97.9% of the original variables. Component-wise ablation showed the largest mean R2 decrease when FB-CARS was replaced; PLSR captured the dominant quantitative structure, and evidence gating selectively admitted residual correction. The complete framework achieved higher target-wise R2 point estimates than the evaluated standalone regressors (mean R2 = 0.8579). Spatial projections of five held-out samples illustrated local PLSR responses without establishing within-sample chemical accuracy. The findings support further evaluation of compact spectral screening, while larger independent datasets are needed before routine quality-control use.