Accurate and early estimation of harvest yields allows decisions to be made both in the field (e.g. green harvesting or irrigation) and in the winery (estimation of wine volume).
The estimation of the harvest yield in advance allows decisions to be made both in the field (e.g. green harvesting or irrigation) and in the winery (estimation of wine volume) and is essential to define the quality of the vintage. For this reason, the field gauging carried out by Pago de Carraovejas (PDC) is extremely laborious, where not only the number of bunches per vine is counted, but also the number of grapes per bunch and the weight of the berry.

GMV had developed a machine-learning (ML) model for estimating gauging over 1300 hectares on the basis of a field gauging reduced to a few plots and five years’ processing of Sentinel-2 satellite data (Copernicus mission). The first test of this model in the PDC vineyards yielded a prediction of 82% accuracy with respect to the volume of grape intake, the starting point of the research.

With the data obtained from the PDC gauging, brought from previous experience, the factor is changed in the theoretical equation. In addition, for each vineyard, we extracted the factor of faults (uprooted vines) from a very high resolution satellite image (50 centimeters). This information has been decisive both in the adjustment of the number of vines in each vineyard from their planting frames and in the adjustment of the historical Sentinel-2 value-added products.

The new ML model developed for PDC has yielded an accuracy of 91% for grapes entering the winery in 2020, rising to 94% in 2021, and has improved the accuracy achieved from field gauging by 2%.