Using Deep Learning to Predict Plant Growth and Yield in Greenhouse Environments

Abstract

Greenhouse growers and farmers in general rely heavily on accurate growth and yield
predictions. Growers may increase environmental control, productivity, supply/demand
matching, and operational expenses with well-modelled growth and yield. Recent advances in
ML, and in particular Deep Learning (DL), can give potent new analytical tools. In this study,
we apply ML and DL methods to estimate production and plant growth variance in two
greenhouse settings: tomato yield forecasting and Ficus benjamina stem growth. In the
prediction formulations, we use a novel deep recurrent neural network (RNN) based on the
Long Short-Term Memory (LSTM) neuron model. In order to model the desired growth
parameters, the RNN design takes into account both the historical yield, growth, and stem
diameter values and the microclimate circumstances. To assess the efficacy of various ML
techniques, such as support vector regression and random forest regression, a comparative
analysis is presented that use the mean square error criterion. Positive findings are given
based on data collected from two greenhouses in Belgium and the United Kingdom as part of
the EU Interreg SMARTGREEN project (2017-2021).