Early yield prediction in different grapevine varieties using computer vision and machine learning
Fernando Palacios, María P. Diago, Pedro Melo‐Pinto, Javier Tardáguila
Abstract
Abstract Yield assessment is a highly relevant task for the wine industry. The goal of this work was to develop a new algorithm for early yield prediction in different grapevine varieties using computer vision and machine learning. Vines from six grapevine ( Vitis vinifera L.) varieties were photographed using a mobile platform in a commercial vineyard at pea-size berry stage. A SegNet architecture was employed to detect the visible berries and canopy features. All features were used to train support vector regression (SVR) models for predicting number of actual berries and yield. Regarding the berries’ detection step, a F1-score average of 0.72 and coefficients of determination (R 2 ) above 0.92 were achieved for all varieties between the number of estimated and the number of actual visible berries. The method yielded average values for root mean squared error (RMSE) of 195 berries, normalized RMSE (NRMSE) of 23.83% and R 2 of 0.79 between the number of estimated and the number of actual berries per vine using the leave-one-out cross validation method. In terms of yield forecast, the correlation between the actual yield and its estimated value yielded R 2 between 0.54 and 0.87 among different varieties and NRMSE between 16.47% and 39.17% while the global model (including all varieties) had a R 2 equal to 0.83 and NRMSE of 29.77%. The number of actual berries and yield per vine can be predicted up to 60 days prior to harvest in several grapevine varieties using the new algorithm.