Development of a predictive modeling framework using hyperspectral images to assess grape biochemical quality

Smart viticulture systems are needed for real‑time, nondestructive assessment of grape biochemical quality to support efficient, modern vineyard management and enable reliable robot‑assisted harvest decisions. This study investigated the use of hyperspectral imaging for biochemical quality assessment of grape berries and bunch-level mold detection. Grape bunches harvested from the Tsurunuma winery vineyard, Hokkaido, Japan, were utilized in this experiment. The experimental setup was a controlled indoor imaging platform with halogen lamps as the illumination source. Hyperspectral signatures of 108 healthy grape berries and 176 mold-affected grape berries were acquired and analyzed to identify discriminative wavelengths predictive of 10 biochemical quality parameters, with particular focus on total soluble solids (TSS), pH, and total acid (TA). Machine learning, deep learning, and hybrid models were trained and benchmarked, and the best-performing models were subsequently integrated with YOLOv11 instance segmentation to collate berry-level spectral predictions into bunch-level quality assessments and percent-good metrics. Spectral analysis of healthy and mold‑affected berries revealed clear reflectance differences (peak deviation 0.039 at 800 nm), with mold‑affected samples showing lower TSS and TA but higher pH. The models achieved strong predictive performance (R² ≈ 0.67–0.97), with class‑conditional modeling reducing systematic bias in TSS and TA. YOLOv11 instance segmentation model more reliably detected mold‑affected than healthy berries (mAP₅₀ = 0.526, precision = 0.475, recall = 0.561). Fusing YOLO detections with hyperspectral predictions enabled bunch‑level percent‑good scoring, which could assist selective harvesting with real‑time robotic operation.