Hyperspectral Imaging and Interpretable Machine Learning for Bleaching-Level Identification and Treatment-Associated SO2 Residue Estimation in Lanzhou Lily

Rapid and nondestructive assessment of sodium metabisulfite (Na2S2O5) bleaching treatments and sulfur dioxide (SO2) residues is important for the quality and safety control of Lanzhou lily. This study integrated hyperspectral imaging (HSI), machine learning, and SHapley Additive exPlanations (SHAP) to identify bleaching levels and estimate treatment-associated SO2 residues. Model performance was assessed using treatment-replicate-grouped validation with samples from a single procurement lot. For bleaching-level classification, the support vector machine (SVM) combined with normalization (NOR), denoted as SVM-NOR, achieved the highest accuracy of 0.902. For SO2 residue estimation, the random forest (RF) model using raw spectra, denoted as RF-Raw, provided the best performance, with a cross-validated coefficient of determination (R2) of 0.978 and a root mean square error (RMSE) of 38.5 mg/kg. The reported regression performance reflects treatment levels represented in calibration. Complementary leave-one-concentration-out (LOCO) validation yielded a higher pooled RMSE of 155.4 mg/kg for held-out levels. SHAP analysis revealed different wavelength-contribution patterns for classification and regression, providing an interpretable basis for the model predictions. These findings suggest the potential of HSI for rapid prescreening, while further validation is needed for predictions between calibrated treatment levels, particularly near the regulatory reference.