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

Understanding grape composition at harvest is vital for winemakers and growers to make informed decisions regarding grape processing and wine quality. The composition of grapes, including sugars (total soluble solids), pH, total acidity (titratable acidity), and anthocyanins, directly influences the flavor, aroma, and quality of the resulting wine. Unfortunately, grape composition varies in space because of environmental and managment factors. Implementing variable rate management can help underperforming areas recover or can help separate areas for selective production of different wines. Unfortunately, current sensing methods map grape composition indirectly. Selective harvesting cannot be implemented in real-time, requiring two separate passes to harvest the same vineyard block, increasing the time and logistic complexity of the harvest.

In this project, a novel approach using hyperspectral imaging technology mounted on the conveyor belt of a mechanical harvester was employed to assess grape composition on the go. The hyperspectral camera captured images within the shortwave infrared (SWIR) domain, ranging from 900 to 1700 nm. This spectral range allows for detecting specific molecular vibrations associated with various chemical compounds present in grapes, such as sugars, acids, and water content.

The images obtained from the hyperspectral camera were processed using advanced segmentation techniques to extract only the grape signal, effectively isolating the relevant spectral information for analysis. By focusing solely on the grape signal, the project aimed to minimize interference from background noise and non-grape elements, ensuring the accuracy of the composition assessment. The extracted grape signal was used to predict grape composition through previously trained machine-learning algorithms and predictions were compared to ground data collected on grape samples from 100 vines. Through this predictive modeling approach, the project seeks to develop a reliable and non-destructive method to assess grape composition in real-time during harvest. This could empower mapping and separate grape composition for on-the-fly, one-pass selective harvesting.

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