Decoupling tea-bud heap structure from non-imaging hyperspectral spectra for accurate single-bud trace biochemistry retrieval

Accurate, real-time, non-destructive estimation of single-bud biochemistry is critical for managing green-tea quality, yet non-imaging hyperspectral measurements of arbitrarily stacked buds are confounded by structured interference from mixed adaxial/abaxial surfaces and their three-dimensional arrangement. We propose Bilinear Spectral Derivative Gaussian Process Regression (BSDGPR), which couples a bilinear spectral-derivative model with Gaussian process regression to minimise structured interference. First- and second-order spectral derivatives attenuate the weakly wavelength-dependent leaf-spatial-arrangement term ( R LSA ), while unit-sphere normalization combined with a chord-distance radial basis function kernel eliminate the plan-view leaf-front area fraction ( f ). This effectively standardizes tea-bud heap spectra to single-bud equivalents. (i) Structure suppression and invariance: independent-contribution analysis shows that moving to the derivative domain boosts the heap term's explanatory power to 80.7–99.4 % across bands while reducing R LSA to ≤9.3 % (TH:LSA ratio of between 4.8 and 40.9), and the sphere-normalized chord-distance kernel cancels f -driven gain variation across samples. (ii) Accuracy and efficiency: relative to reflectance-domain GPR, BSDGPR raises R 2 by 0.13 and lowers NRMSE by 3–5 percentage points for tea polyphenols (0.68 → 0.81; 15.55 % → 12.64 %), amino acids (0.56 → 0.71; 19.78 % → 15.72 %), and soluble sugars (0.86 → 0.91; 10.85 % → 9.05 %), while maintaining fast inference (10–13 s per model) suitable for near real-time use. (iii) Transferability: models trained on tea-bud heaps generalize to canopy measurements with minimal loss (ΔR 2 ≤ 0.05; canopy NRMSE 11–15 %), remain credible across cultivars (R 2 = 0.37–0.74; NRMSE ≤23 % across 15 varieties), are stable at typical sensor resolutions (≤ 10–15 nm), and achieve R 2 = 0.63–0.66 with NRMSE ≤22 % for wheat canopy chlorophyll. Overall, BSDGPR is a reliable, field-ready AI approach for non-destructive biochemical sensing across crops, sensors, and management scenarios, offering broad applicability and strong potential for precision agriculture. • Physically interpretable BSDGPR converts heap spectra to single-bud equivalents, suppressing R LSA and f . • BSDGPR exceeds GPR accuracy, raising R 2 by 0.13 and lowering NRMSE by 3–5 % across constituents. • Inference is near real time, requiring approximately 10–13 s per model on field spectra. • Models transfer robustly from heaps to canopies (ΔR 2 ≤ 0.05) and across cultivars (NRMSE ≤23 %).