A Stride Toward Wine Yield Estimation from Images: Metrological Validation of Grape Berry Number, Radius, and Volume Estimation

Abstract

Yield estimation is a key point theme for precision agriculture, especially for small fruits and in-field scenarios. This paper focuses on the metrological validation of a novel deep-learning model that robustly estimates both the number and the radii of grape berries in vineyards using color images, allowing the computation of the visible (and total) volume of grape clusters, which is necessary to reach the ultimate goal of estimating yield production. The proposed algorithm is validated by analyzing its performance on a custom dataset. The number of berries, their mean radius, and the grape cluster volume are converted to millimeters and compared to reference values obtained through manual measurements. The validation experiment also analyzes the uncertainties of the parameters. Results show that the algorithm can reliably estimate the number (MPE=−5%, σ=6%) and the radius of the visible portion of the grape clusters (MPE=0.8%, σ=7%). Instead, the volume estimated in px3 results in a MPE=0.4% with σ=21%, thus the corresponding volume in mm3 is affected by high uncertainty. This analysis highlighted that half of the total uncertainty on the volume is due to the camera–object distance d and parameter R used to take into account the proportion of visible grapes with respect to the total grapes in the grape cluster. This issue is mostly due to the absence of a reliable depth measure between the camera and the grapes, which could be overcome by using depth sensors in combination with color images. Despite being preliminary, the results prove that the model and the metrological analysis are a remarkable advancement toward a reliable approach for directly estimating yield from 2D pictures in the field.

Keywords

• measurement science
• machine vision
• viticulture
• deep learning
• fruit counting
• fruit size estimation