Autonomous Predictions of Vineyard Yield Through Machine-Learning Modeling of Remote-Sensing and Historical Data
*California State University Fresno, 2360 E Barstow Avenue, Fresno, CA 93740 (firstname.lastname@example.org)
This work presents a very novel approach to predict vineyard yield at the block level from machine-learning analysis of remote-sensing and historical yield data in a fully autonomous way. The work was performed on a 200-acre ranch in California, divided into 16 blocks. Fifteen years of yield data were available, since the first year of production for the ranch. The ranch was planted with multiple varieties (Cabernet Sauvignon, Merlot, and Cabernet franc), but the rootstock was always 1103P. Even within the same year, variability in yield/block was very large, ranging from ~2 to 10 tons. Weather and soil conditions were also very variable considering the time span of the dataset, the surface of the ranch, and the hillslope conditions. Weather data were obtained from the closest weather station of the California Irrigation and Management System (CIMIS), and soil data obtained from the USDA soil database (SSURGO). Terrain analysis was performed using a digital elevation model to obtain the soil wetness index, aspect, and slope of each vineyard block. Over 600 Landsat 7 images for all years were cloud-filtered, SCRI-off imputed, and atmospherically and topographically corrected. Over 15 different multispectral indexes were computed and integrated over time for all vineyard blocks. A machine-learning meta-ensemble of random forest and extreme gradient boosting was then trained to estimate block yield from this data set of static and dynamic data. The machine was subjected to multiple validation routines and was finally able to estimate vineyard yield by the month of July with an error of 0.72 tons/acre, or less than 15%, in a three-year test set of unobserved data. Interpretation of the model provides new insights about environmental causes of yield variations in space and time. This is a very novel approach that, included in a software-as-a-service for growers, promises to change yield estimation practices in the grape industry.
Funding Support: California State University – startup funds