Mapping grape composition in the field using VIS/SWIR hyperspectral cameras mounted on a UTV
Context and purpose of the study – Assessing grape composition is critical in vineyard management. It is required to decide the harvest date and to optimize cultural practices toward the achievement of production goals. The grape composition is variable in time and space, as it is affected by the ripening process and depends on soil and climate conditions. This variability makes an appropriate assessment of the overall grape composition of a vineyard block complicated and time-consuming. Our work focused on developing a system to assess and map grape composition directly in the field through the application of machine-vision models to hyperspectral images acquired on the go in a vineyard with sprawling canopies, where the fruit tends to be hidden by the foliage.
Material and methods – For this study, a UTV was specially adapted to lift the canopy and expose the fruits, two hyperspectral cameras (a Senop HSC VIS/NIR and a Specim NIR/SWIR) were mounted with GPS systems and halogen lights for night imaging. We imaged a Merlot vineyard located in Madera, California, four times during the 2022 growing season. At the same time, we sampled grapes from 160 vine locations which were analyzed in the laboratory to assess anthocyanin, soluble solids, pH, and titratable acidity. A total of ~1,000 samples were collected. For the analysis, the images needed to be segmented to extract the grape’s signal from sampled vines. Then, the reflectance of the grapes was used to look for correlations with grape composition using machine learning models. Evaluation and interpretation of models were performed using RMSE, R2. Interpretation of the model was conducted through feature importance and partial dependence plots to understand the relationship between wavelength predictors and the outcome. This project is the first to use a SWIR camera mounted on a UTV to assess grape composition.
Results – Our results demonstrate that SWIR images can be used to perform a classification to extract grape signal with a mean error of 2.2% using the spectral signature of each class represented in the image (grape, leaves and background). The prediction of grape compounds from the refined spectral signal shows promising results. This project aims to help growers to monitor grape composition in the field rapidly and spatially to inform variable rate management.
Issue: GiESCO 2023
1Department of Viticulture & Enology, California State University Fresno, Fresno, CA, USA
2Viticulture and Enology Research Center, California State University, Fresno CA, USA
3Winegrowing Research Department, E&J Gallo Winery, Modesto, CA, USA
Contact the author*
precision viticulture, grape composition, hyperspectral imaging, mapping, machine-learning