Grape Yield Prediction Models: Approaching Different Machine Learning Algorithms

Abstract

Efficient marketing of winegrapes involves negotiating with potential buyers long before the harvest, when little is known about the expected vintage. Grapevine physiology is affected by weather conditions as well as by soil properties and such information can be applied to build yield prediction models. In this study, Partial Least Squares Regression (PLSR), Cubist (CUB) and Random Forest (RF) algorithms were used to predict yield from imputed weather station data and soil sample analysis reports. Models using only soil variables had the worst general results (R2 = 0.15, RMSE = 4.16 Mg ha−1, MAE = 3.20 Mg ha−1), while the use of only weather variables yielded the best performance (R2 = 0.52, RMSE = 2.99 Mg ha−1, MAE = 2.43 Mg ha−1). Models built with CUB and RF algorithms showed signs of overfitting, yet RF models achieved the best average results (R2 = 0.58, RMSE = 2.85 Mg ha−1, MAE = 2.24 Mg ha−1) using only weather variables as predictors. Weather data imputation affected RF and CUB models more intensely while PLSR remained fairly insensitive. Plant age, yield level group, vineyard plot, May temperatures, soil pH and exchangeable concentrations of Zn, Cu, K and Mn were identified as important predictors. This exploratory work offers insights for future research on grape yield predictive modeling and grouping strategies to obtain more assertive results, thus contributing to a more efficient grapevine production chain in southern Brazil and worldwide.