Automated Development of Model to Predict Tea Harvest Date from Meteorological Data Collected on Site

Abstract

The time of picking is crucial to the production of green tea: delays reduce quality, but picking early reduces yield. Therefore, it is necessary to pick tea at the optimum time to balance quality and yield. At present, deciding when to pick relies on individual experience. However, climate change could make experience less reliable. Therefore, objective prediction of tea harvest dates will be important, not only for stabilizing the production of high-quality tea, but also for saving labor and management. Here, we report multiple linear regression models automatically constructed for predicting the harvest date of cultivars Yabukita, Sayamakaori, and Okumidori in Japan through the use of meteorological data collected over many years on site. The aim was to limit the mean absolute error of models evaluated by the leave-one-out strategy to within 3 days. All models performed better than simple linear regressions. The techniques developed here will reduce the effort of visual judgment and contribute to the stable production of high-quality tea.