Machine-Learning Methods to Identify Key Predictors of Site-Specific Vineyard Yield and Vine Size

Abstract

Background and goals Lake Erie Concord growers have access to high-resolution spatial soil and production data, but lack protocols and information on the optimum time to collect these data. This study examines the type and timing of sensor information to support in-season management. Methods and key findings A three-year study in a 2.6 ha vineyard collected yield, pruning mass, canopy vigor, and soil data, including yield and pruning mass from the previous year, at 321 sites. Stepwise linear regression and random forest regression approaches were used to model site-specific yield and pruning mass using historical spatial production data, multi-temporal in-season canopy vigor, and soil data. The more complex yield elaboration pro-cess was best modelled with non-linear random forest regression, while the simpler development of pruning mass was best modelled by linear regression. Conclusions and significance Canopy vigor in the weeks preceding bloom was the most important predictor of the current season’s yield and should be used to generate stratified sampling designs for crop estimation at 30 days after bloom. In contrast, pruning mass was not well-predicted by canopy vigor; even late-season canopy vigor, which is widely advo-cated to estimate pruning mass in viticulture. The previous year’s pruning mass was the dominant predictor of pruning mass in the current season. To model pruning mass going forward, the best approach is to start measuring it. Further work is still needed to develop robust, local site-specific yield and pruning mass models for operational decision-making in Concord vineyards.