Accurate estimation of grapevine bunch weight using image analysis: a case study with two Portuguese cultivars

Vineyard yield estimation can bring several benefits to all the grape and wine production chain. Among several methods the ones based on estimation of yield components are the most used at farm level. However, as they are manual, destructive and very time-consuming, there is a strong demand to replace them with low-cost and reliable automated methods. Recent advances in machine vision have provided accurate tools for bunch and/or berry recognition. However, converting the visible bunch area on the images into bunch mass is still a big challenge. In the frame of the EU VINBOT research project (www.vinbot.eu), an experiment was set up using the cultivars Viosinho (white; 100 bunches) and ‘Trincadeira’ (red, 48 bunches) to study the relationships between the projected bunch area (Ba) on the 2D images and the corresponding bunch weight (Bw) measured at harvest. In the laboratory, bunches were submitted to image acquisition using a compact RGB camera. Then each bunch was assessed to obtain the following morphological attributes: Bw, bunch volume (Bv), berry number (BE#) and weight (BEw) and rachis length (Rl). Bunch compactness (Bc) was calculated as the ratio between BE# and Rl, while the Ba was computed using ImageJ^® software. Correlation analysis shows that most part of these variables are significantly and positively correlated with Bw. However, as not all variables are easy to obtain by automated image analysis, some were excluded and a forward stepwise regression between Bw (dependent variable) and the variables BE#, Ba, Bv and Bc (independent variables) was performed. The final models obtained explained a very high proportion of bunch weight variability (R²=0.98 and 0.99 for ‘Viosinho’ and ‘Trincadeira’, respectively) with a very small error. These results indicate that grapevine bunch weight can be estimated with high accuracy from 2D images using explanatory variables derived from bunch morphological attributes.