The effect of cultivar on the conversion of grape pixels into yield at grapevine level

Vineyard yield estimation is a challenge that has been subject to research for several years. Common practices are invasive, labour intensive and often inaccurate, as vineyard yield varies both spatially and temporally. Non-invasive, image-based grapevine yield estimation methods focus on automatically identifying yield components in digital images which are then converted into yield. Such systems are dependent on the accuracy of yield components detection in the images and corresponding conversion into mass. Several works mention the potential of bunch pixels (projected area; BA) as an estimator of bunch weight. However, most relationships have been studied in laboratory conditions and/or on single bunches. At plant level, from a 2D perspective, such relationships depend on occlusions by other bunches, as well as bunch morphology. The present work aims at comparing the relationships between actual grapevine total bunch weight and corresponding BA on four cultivars, in order to estimate grapevine yield from 2D images taken in field conditions. The study included four cultivars (‘Arinto’, ‘Encruzado’, ‘Castelão’, ‘Syrah’) from two commercial vineyards, trained on a vertical shoot positioning system, located at two different sites within the Lisbon winegrowing region, Portugal. Images were collected from 100 vines defoliated at bunch zone and BA was segmented manually from each image. The regression analysis between BA and actual yield presented high and significant R² for all cases (>0.90), however, some significant slope differences were detected among fitted lines, indicating cultivar dependency. No significant slope differences were found between regression lines from pooled data of the two sites. Our work corroborates recent research regarding the use of BA as a good estimator of bunch weight, however, in a multi-cultivar scenario, this relationship should be considered with caution. Further research is ongoing in order to create an accurate generalized model using other image-based complementary variables.