Grapevine bunch weight estimation using image-based features: comparing the predictive performance of number of visible berries and bunch area

Abstract

Recent advances in machine vision technologies have provided a multitude of automatic tools for recognition and quantitative estimation of grapevine bunch features in 2D images. However, converting them into bunch weight (BuW) is still a big challenge. This paper aims to compare the explanatory power of the number of visible berries (#vBe) and the bunch area (BuA) in 2D images, in order to predict BuW. A set of 300 bunches from four grapevine cultivars were picked at harvest and imaged using a digital RGB camera. Then each bunch was manually assessed for several morphological attributes and, from each image, the #vBe was visually assessed while BuA was segmented using manual labelling combined with an image processing software. Single and multiple regression analysis between BuW and the image-based variables were performed and the obtained regression models were subsequently validated with two independent datasets.The high goodness of fit obtained for all the linear regression models indicates that either one of the image-based variables can be used as an accurate proxy of actual bunch weight and that a general model is also suitable. The comparison of the explanatory power of the two image-based attributes for predicting bunch weight showed that the models based on the predictor #vBe had a slightly lower coefficient of determination (R2) than the models based on BuA. The combination of the two image-based explanatory variables in a multiple regression model produced predictor models with similar or noticeably higher R2 than those obtained for single-predictor models. However, adding a second variable produced a higher and more generalised gain in accuracy for the simple regression models based on the predictor #vBe than for the models based on BuA. Our results recommend the use of the models based on the two image-based variables, as they were generally more accurate and robust than the single variable models. When the gains in accuracy produced by adding a second image-based feature are small, the option of using only a single predictor can be chosen; in such a case, our results indicate that BuA would be a more accurate and less cultivar-dependent option than the #vBe.