Reaction-informed spectral feature packets for lightweight Vis/NIR screening of sulfur-fumigation process states in plant-derived materials

Sulfur fumigation induces ordered optical and matrix-related changes in plant-derived materials, including bleaching-like reflectance variation, moisture/matrix alteration, and spatial heterogeneity. This study developed a lightweight Sulfur Response-informed Spectral Feature Packet (SR-SFP) framework for rapid screening of predefined fumigation process states. Chemical reference measurements were incorporated to evaluate the correspondence between the process-state labels and fumigation-associated compositional variation. Two 31-band Vis/NIR imaging datasets were analyzed: Fructus Aurantii (400 full-ROI samples) and processed Pinelliae Rhizoma (800 full-ROI samples), each covering untreated control, light, moderate, and heavy fumigation states. SR-SFP transformed ROI spectra and image-derived heterogeneity descriptors into interpretable packets representing bleaching-like response, moisture/matrix response, tissue heterogeneity, and ordinal process-state evolution. Plate-grouped validation prevented ROI leakage, and all preprocessing and feature-selection operations were fitted exclusively within training folds. The best conventional benchmark, SNV+D1-XGBoost, achieved a mean Macro F1 of 0.9238, whereas SR-SFP-XGBoost achieved 0.9113 with comparable severe-error control. Ablation and 64-feature compression confirmed robust lightweight screening performance. SR-SFP therefore provides an interpretable strategy for fumigation process-state screening and risk grading.