Estimating the trajectories of vine cordons in full foliage canopies for automated green shoot thinning in vineyards

Abstract Green shoot thinning in vineyards is carried out annually aiming to produce high-quality grapes. Recently, mechanical machines are used for green shoot thinning to reduce labor costs which causes 10–85% variation in shoot removal efficiencies due to the difficulty in precisely positioning the thinning end-effector to cordon trajectories. Automatically controlling the position and orientation of thinning end-effector to follow the cordon trajectories is expected to increase the performance and efficiency of these machines. However, during most of the green shoot thinning season (between first week of bud opening to the fourth week of shoot growth), cordons are occluded by shoots/leaves making it extremely challenging to determine the trajectories of cordons. In this study, we presented a deep learning based novel approach to estimate the trajectories of cordons during the thinning season in real field conditions. First, deep residual network (ResNet) and Faster region-based convolutional neural network was used to detect visible segments of full foliage grapevine canopies. Then, location information of detected visible segments were used to estimate the trajectories of cordons. This approach was evaluated using ground truth images collected from different growth stages of green shoots in real vineyard environment. The proposed approach achieved a correlation coefficient of 0.993, 0.991, and 0.987 in estimating cordon trajectories with the test dataset from weeks 2 to 4, respectively, of shoot growth.