Using a Real-Time Photosynthesis and Transpiration Monitoring System to Develop Random Forests Models for Predicting Cherry Tomato Yield in a Commercial Greenhouse

Abstract

　Predicting the yield of horticultural crops is crucial to meet the expectations of retailers and consumers. In this study, we developed random forests (RF) based on the measured amounts of whole-plant photosynthesis and transpiration to predict cherry tomato fruit yields in a commercial greenhouse in Japan. Whole-plant daily net photosynthesis (Photo) and daily transpiration (Trans) were measured by using a real-time photosynthesis and transpiration monitoring system. Variables of environmental conditions (Env), including daily solar irradiation, air temperature, and atmospheric water vapor deficit, were also measured in the greenhouse. Data with different 7 variable combinations (Env, Photo, Trans, Env+Photo, Env+Trans, Photo+Trans, Env+Photo+Trans) and different 21 timeframes (from 1 to 6 consecutive weeks in the past 6 weeks) were used to train models for predicting the yield for the subsequent week. RF models with the timeframes of 3 consecutive weeks until 2 weeks before the date of yield prediction (3W2) and 4W2 and variable combinations of Photo, Env+Photo, and Photo+Trans had relatively low normalized root mean square error (RMSE%; 9.8-10.3%). The model that had a timeframe 4W2 and variable combination Photo had the best accuracy (RMSE% ＝ 9.8%). These indicate that whole-plant photosynthesis and transpiration are good predictors of cherry tomato yield.