On-the-go Mapping of Grape Composition in the Field Through Hyperspectral Machine-Vision

This project reports the first six months of activity for this project. The goals are to develop predictive models to assess grape composition (total soluble solids, titratable acidity, anthocyanins) from hyperspectral imaging of berries and identify the most important wavelengths to foster the development of simplified sensors specifically adapted to monitor grapes in the field with reduced complexity and cost. It also plans to develop methods to deploy the sensors in the field, map grape composition on the go, and inform variable rate strategies to improve grape composition with spatially tailored canopy management techniques. We used the varietal collection at Fresno State to develop a spectral library of grape berries, including all commodities, wine grapes, table grapes, juice grapes, and raisins. We collected 496 samples from 77 black, red, and white varieties and nine sampling dates throughout the ripening. We imaged them with a VIS-NIR hyperspectral camera in controlled light conditions in the laboratory. We analyzed samples to measure °Brix, pH, titratable acidity, and the anthocyanin profile through HPLC. As planned, we are now developing machine-learning models to predict the sample composition. To field-test the system, we have identified two vineyards located in Madera and Fresno county and easy to reach from our campus. The vineyards show variability in grape composition at a short scale. We have been characterizing this variability with a randomized sampling based on the yield map at harvest. We have acquired imagery in the field and can segment clusters from grape canopies; we are now optimizing the sensing support system to obtain high-quality hyperspectral images in the field.