Articles
Tomato yield prediction in a semi-closed greenhouse
Article number
1107_36
Pages
263 – 270
Language
English
Abstract
Tomato yields often vary from week to week, so the ability to accurately predict them would give producers a competitive advantage.
As a result there has been interest from growers and researchers about tomato yield prediction.
The purpose of this study is to predict tomato yield in a semi-closed and reference experimental greenhouses, located at Humboldt University of Berlin.
At this location a range of physiological as well as environmental variables have been automatically monitored.
A Dynamic Artificial Neural Network (DANN) was implemented to predict yields in both greenhouses for comparison purposes.
The input data were CO2 fixation, transpiration, and solar radiation, as well as past yield.
The output is the yield for the next week.
Different networks were built and evaluated to suggest a certain network structure that provides optimal results for this particular case.
A sensitivity analysis was conducted to estimate the relative importance of the different inputs to model predictions.
The most important variable for yield prediction was CO2 fixation, and the least important was transpiration.
Given the high costs of the phytomonitoring systems, our interest was to prove that yield can be predicted using only external conditions and yield with delays, thus using only solar radiation as the input variable gave a correlation coefficient of R=0.9165 for training validation and testing, R=0.9784 for simulation, the absolute error in both cases was very small.
These results show the DANN is as a powerful tool for prediction with few data patterns.
As a result there has been interest from growers and researchers about tomato yield prediction.
The purpose of this study is to predict tomato yield in a semi-closed and reference experimental greenhouses, located at Humboldt University of Berlin.
At this location a range of physiological as well as environmental variables have been automatically monitored.
A Dynamic Artificial Neural Network (DANN) was implemented to predict yields in both greenhouses for comparison purposes.
The input data were CO2 fixation, transpiration, and solar radiation, as well as past yield.
The output is the yield for the next week.
Different networks were built and evaluated to suggest a certain network structure that provides optimal results for this particular case.
A sensitivity analysis was conducted to estimate the relative importance of the different inputs to model predictions.
The most important variable for yield prediction was CO2 fixation, and the least important was transpiration.
Given the high costs of the phytomonitoring systems, our interest was to prove that yield can be predicted using only external conditions and yield with delays, thus using only solar radiation as the input variable gave a correlation coefficient of R=0.9165 for training validation and testing, R=0.9784 for simulation, the absolute error in both cases was very small.
These results show the DANN is as a powerful tool for prediction with few data patterns.
Authors
R. Salazar, I. López, A. Rojano, U. Schmidt, D. Dannehl
Keywords
dynamic neural networks, transpiration, CO2 fixation, solar radiation
Online Articles (43)