Hongrui Wang and Jason Londo*
*Cornell University, 635 W North St, Geneva, NY, 14456 (jpl275@cornell.edu)

Accurate, real-time monitoring of grape freezing tolerance is crucial for the sustainability of the grape industry in cool climate viticultural regions. However, on-site measurement is limited due to the complexity of measuring method. Current prediction models, generated by parameter-based modeling using local measurement data, underperform under other climate conditions, preventing large-scale deployment of these methods. Here, we combined grape freezing tolerance measurement data from multiple sites in North America and generated a prediction model based on newly developed hourly temperature-derived features and one-hot encoded cultivar features using AutoGluon, an automatic machine learning engine. The final model was tested and compared with previous biological models for performance under different climate conditions. Feature importance of the model was quantified by computing SHAP value. The final model (weighted ensemble model level 2) achieved an overall 1.31°C root-mean-square error during model testing. The model also out-performed the current standard models at all testing sites. Finally, the model was deployed in the major viticultural regions in the Northeastern and Midwestern United States, using daily-updated weather data in Applied Climate Information System (ACIS). A real-time freezing tolerance and freezing damage prediction system was developed and launched using R shiny.

Funding Support: Cornell University CALS School of Intergrative Plant Science Horticulture Section Extension Outreach Assistantship