Articles
Time-series prediction of greenhouse environmental conditions and plant growth indicators using RNN-based models
Article number
1423_43
Pages
343 – 349
Language
English
Abstract
The primary goal of greenhouse environmental control is to create an optimal microclimate that promotes plant growth and optimizes production.
Predicting greenhouse conditions improves management strategies that can lead to increased crop yields with fewer resources.
The objective of this study is to predict greenhouse environmental conditions (temperature, relative humidity, and CO2 concentration) and plant growth indicators (stem elongation, stem thickness, and cumulative number of trusses) using two recurrent neural network-based models: long short-term memory (LSTM) and segmented recurrent neural networks (SegRNN). The study evaluates the prediction performance of these models using two metrics: root mean squared error (RMSE) and mean absolute error (MAE). The predictions of the environmental changes after 2 and 24 h, as well as the plant growth indicators after 1 and 7 days, were carried out using data from a 5-month cultivation of cherry tomatoes in a greenhouse.
The results show that the SegRNN outperformed the LSTM in predicting greenhouse conditions.
Specially, for the prediction of CO2 after 2 and 24 h, SegRNN achieved MAE values 0.357 and 0.630, respectively; while LSTM predicted lower MAE values of 0.885 and 1.002, respectively.
Predicting plant growth metrics after 1 and 7 days, SegRNN showed better performance with MAE values of 0.007 and 0.027, respectively, for cumulative number of trusses.
In contrast, LSTM showed poor performance with considerably lower MAE values of 0.712 and 1.199 for these growth metrics.
Overall, SegRNN showed satisfactory performance even with the limited growth data available in this study.
Predicting greenhouse conditions improves management strategies that can lead to increased crop yields with fewer resources.
The objective of this study is to predict greenhouse environmental conditions (temperature, relative humidity, and CO2 concentration) and plant growth indicators (stem elongation, stem thickness, and cumulative number of trusses) using two recurrent neural network-based models: long short-term memory (LSTM) and segmented recurrent neural networks (SegRNN). The study evaluates the prediction performance of these models using two metrics: root mean squared error (RMSE) and mean absolute error (MAE). The predictions of the environmental changes after 2 and 24 h, as well as the plant growth indicators after 1 and 7 days, were carried out using data from a 5-month cultivation of cherry tomatoes in a greenhouse.
The results show that the SegRNN outperformed the LSTM in predicting greenhouse conditions.
Specially, for the prediction of CO2 after 2 and 24 h, SegRNN achieved MAE values 0.357 and 0.630, respectively; while LSTM predicted lower MAE values of 0.885 and 1.002, respectively.
Predicting plant growth metrics after 1 and 7 days, SegRNN showed better performance with MAE values of 0.007 and 0.027, respectively, for cumulative number of trusses.
In contrast, LSTM showed poor performance with considerably lower MAE values of 0.712 and 1.199 for these growth metrics.
Overall, SegRNN showed satisfactory performance even with the limited growth data available in this study.
Authors
Y.L. Kim, H.J. Park, D.Y. Kim, J.H. Cho, D.S. Shin, D.H. Lee, S.H. Park, H.K. Suh
Keywords
time-series prediction, greenhouse environment prediction, plant growth prediction, lstm, segrnn
Online Articles (43)