Deep learning and computer vision for assessing the number of actual berries in commercial vineyards

The number of berries is one the most relevant yield components that drives grape production in viticulture. goal this work was to estimate actual per grapevine using computer vision and deep learning commercial vineyards. Images from visible range (RGB) were acquired a set 96 grapevines (Vitis vinifera L.) at pea-size berry stage red, green blue camera (RGB). At harvest, vine manually assessed as ground-truth values. algorithm involved detect images extract canopy features, order gain information about occlusion. These used by machine regression models built vine. A SegNet architecture segment individual several related features. Four datasets created combining estimated different Three tested on four datasets. best results achieved with support vector (SVR) dataset including six This method yielded root mean squared error (RMSE) 205 berries, normalised (NRMSE) 24.99% coefficient determination (R2) 0.83 between show can be high accuracy up 60 days prior developed based learning.