Image based vineyard yield prediction using empirical models to estimate bunch occlusion by leaves
Context and purpose of the study – Vineyard yield estimation brings several advantages to the entire wine industry. It can provide useful information to support decision making regarding bunch thinning practices, harvest logistics and marketing strategies, as well as to manage stored wine and cellar tanks allocation. Today, this estimation is performed mainly using manual methods based on destructive bunch sampling. Yield estimation using image analysis has the potential to perform this task extensively, automatically and non-invasively. However, bunch occlusion, caused mainly by leaves, presents a great challenge to this approach. This occlusion is highly dependent on canopy porosity, which in turn is affected by factors such as vigor, shoot density and leaf area, water availability, biotic and abiotic stresses, among others. In this work, the results of an image-based yield estimation method that estimates bunch occlusion by leaves using canopy porosity data, are compared with a manual approach.
Material and methods – The trial was carried out in two vineyards located within Lisbon winegrowing region, over four years (2018-21). Spur pruned vines trained on a vertical shoot positioning trellis system were used. In a first step, an empirical model was computed to estimate the fraction of bunches occluded by leaves based on the proven assumption in the literature that there is a relationship between canopy porosity and the fraction of exposed bunches. For this, images were captured from 1 m segments at two phenological stages (veraison and full maturation) in non-defoliated and partially defoliated vines of three grape varieties. This model was then used, in a second step, along with other image-based predictors of bunch weight, to estimate grapevine yield. The developed approach included image-based variables related to the visible bunch area and perimeter, berry number and bunch compactness, while considering canopy porosity to estimate the fraction of occluded bunch area. Results were compared to a manual method based on bunch counts and historical bunch weight, on six grape varieties, at veraison. All vine images were collected from a perspective perpendicular to the vine rows, by a static commercial RGB camera or a RGB camera installed on a terrestrial robot.
Results – The yield estimated with the developed algorithm showed a high correlation with the actual yield (R2 = 0.86), with estimation errors ranging between -0.1% and 20.8%, depending on the variety and the year. In most cases, the proposed algorithm outperformed the manual method which was mostly impaired by variations of bunch weight that were not considered by historical data. The proposed image-based approach seems to be an accurate alternative to conventional yield estimation methods. It can be carried out using different image collection setups and has the advantage of being independent of historical data and able to be applied to much larger samples than those used in manual methods. Even though the occlusion estimation method worked well for most cases, further research is needed for modeling non-visible bunches in very dense canopies.
Issue: GiESCO 2023
Linking Landscape, Environment, Agriculture and Food (LEAF), Instituto Superior de Agronomia, Universidade de Lisboa, Tapada da Ajuda, 1349-017 Lisboa, Portugal
Contact the author*
grapevine yield prediction, bunch occlusion, proximal sensing, canopy porosity, bunch pixels