Luca Brillante,* Abiodun Abioye, Eve Laroche-Pinel, Guadalupe Partida, 
Benjamin Corales, Kaylah Vasquez, Vincenzo Cianciola, and Brent Sams
*Department of Viticulture and Enology, California State University Fresno, 
2360 E Barstow Ave, Fresno, CA, 93740 (lucabrillante@csufresno.edu)

Different factors can affect berry composition, such as soil characteristics, water availability, or other environmental factors. Knowing the chemical and physical variability of grapes in a vineyard can help manage growing conditions, avoid over-ripening berries, and plan the harvest. A UTV was adapted especially for this study to help lift the canopy and expose fruits. One hyperspectral camera in the NIR/SWIR domains was mounted on the back with GPS systems and halogen lights for night imaging. With this system, a Merlot vineyard located in Madera, California was imaged four times during the growing season and grapes were sampled for analysis in the laboratory. About 650 samples were collected and georeferenced. The grape signal was extracted through semantic segmentation and separated into grape, leaves, and background. The grape composition was predicted using four models: random forest (RF), extra tree regressor (ETR), extreme gradient boosting (XGboost), and gradient boosting (GB). The model’s performances were assessed using 10-fold cross-validation and an external test set collected in a different vineyard and growing season. Predicting grape composition using the reflectance spectrum exhibited promising results, with ETR having the best performance for the prediction of total soluble solids (R2 = 0.91; NRMSE = 7%), pH (R2 = 0.90; NRMSE = 6 %), titratable acidity (R2 = 0.85; NRMSE = 7.5 %), and total anthocyanins (R2 = 0.91; NRMSE = 7 %). The models were used to develop prediction maps to understand the spatial variability of grape composition attributes in the vineyard and ripening heterogeneity. This study successfully proposes a system to accurately and rapidly monitor grape ripening in the field based on on-the-go hyperspectral mapping.

Funding Support: American Vineyard Foundation, California State University – Agriculture Research Institute