Grapevine Variety Discrimination Using Airborne Hyperspectral Imagery

Grape is the most valuable fruit crop in the United States worth an estimated $5 billion. One of the most pressing issues facing the viticulture industry are grapevine diseases, which can cause total crop and vine loss when left unchecked. An important aspect of grape disease detection is knowledge of grapevine variety, as different varieties have differing susceptibility profiles. Additionally, variety is known to impact spectroscopic disease detection accuracy. Here, we use random forest discrimination to distinguish between 24 major varieties of grapevine based on spectroscopic profile. Training data was compiled from hyperspectral imagery collected by the Advanced Visual/Infrared Imaging Spectrometer Next-Generation (AVIRIS-NG) over Lodi, California in Fall of 2020. Grapevine variety identity was contributed by wine industry collaborators. Data was pre-processed using Sci-Kit Learns k-means clustering algorithm to determine the high-density vegetation pixels within a vineyard plot, in order to minimize the effect of mix-pixel types on the model. These green dense-vegetation pixels were then used to construct a random forest classifier, which exhibited an accuracy of 72.7% and a kappa score of 0.68. Individual variety classification accuracy ranged from 36% to 95%. Future work will seek to improve this model, by both examining systematic misclassifications in the context of grape hybrid parentage, and exploring a spatial random forest to leverage variety heterogeneity within agricultural plots. This work lays the foundation to develop more robust methods of grape variety discrimination, and consequently more advanced methods of disease detection. Alongside expanded availability of hyperspectral datasets in the coming decade (such as those to be provided by NASAs Surface Biology and Geology satellite), a framework for grape variety discrimination and disease detection will serve as a cost effective and sustainable tool for viticultural development.