Machine Learning predictive model of grapevine yield based on agroclimatic patterns

Abstract

Grapevine yield prediction during phenostage and particularly, before harvest is highly significant as advanced forecasting could be a great value for superior grapevine management. The main contribution of the current study is to develop predictive model for each phenology that predicts yield during growing stages of grapevine and to identify highly relevant predictive variables. Current study uses climatic conditions, grapevine yield, phenological dates, fertilizer information, soil analysis and maturation index data to construct the relational dataset. After words, we use several approaches to pre-process the data to put it into tabular format. For instance, generalization of climatic variables using phenological dates. Random Forest, LASSO and Elasticnet in generalized linear models, and Spikeslab are feature selection embedded methods which are used to overcome dataset dimensionality issue. We used 10-fold cross validation to evaluate predictive model by partitioning the dataset into training set to train the model and test set to evaluate it by calculating Root Mean Squared Error (RMSE) and Relative Root Mean Squared Error (RRMSE). Results of the study show that rf_PF, rf_PC and rf_MH are optimal models for flowering (PF), colouring (PC) and harvest (MH) phenology respectively which estimate 1484.5, 1504.2 and 1459.4 (Kg/ha) low RMSE and 24.6%, 24.9% and 24.2% RRMSE, respectively as compared to other models. These models also identify some derived climatic variables as major variables for grapevine yield prediction. The reliability and early-indication ability of these forecast models justify their use by institutions and economists in decision making, adoption of technical improvements, and fraud detection.