Articles
A dynamic artificial neural network for tomato yield prediction
Article number
1154_11
Pages
83 – 90
Language
English
Abstract
A dynamic artificial neural network was developed to predict tomato yield in two high-technology greenhouses located at Humboldt University of Berlin.
One is a reference greenhouse, the other is a semi-closed greenhouse, called a solar collector.
Both greenhouses have a surface area of 307 m2, and they are equipped with a phytomonitoring system for measuring physiological variables.
In addition, the facility has a meteorological station for measuring climatological variables.
Accumulative weekly solar radiation, transpiration and CO2 fixation were the input variables and tomato yield the output variable.
Tomato ‘Pannovy’ yields were recorded from calendar week 17 to 40; data from weeks 17-32 were used for training, validation and testing the neural network, and weeks 33-40 were used to test the ability of the model as a yield predictor.
A layered digital neural network (LDDN), a type of dynamic neural network, was applied, which includes delay lines between the layers, so the output also depends on previous inputs and/or previous states of the network.
The best LDDN model was chosen on the basis of a suitable architecture (i.e., a minimum number of hidden layers and neurons connected to an output layer). Also, two statistical measures were used to examine the goodness of fit between measured and predicted yields, the correlation coefficient (R) and the mean absolute error (MAE). Results show that an LDDN with one hidden layer with four nodes, six delays for the inputs and four delays for the output variable gives the best performance.
The most influencing variable for tomato yield prediction was CO2 fixation, and the less important variable was transpiration.
This work shows the power of the LDDN for prediction purposes, which can be used to predict yields in both greenhouses one week in advance.
One is a reference greenhouse, the other is a semi-closed greenhouse, called a solar collector.
Both greenhouses have a surface area of 307 m2, and they are equipped with a phytomonitoring system for measuring physiological variables.
In addition, the facility has a meteorological station for measuring climatological variables.
Accumulative weekly solar radiation, transpiration and CO2 fixation were the input variables and tomato yield the output variable.
Tomato ‘Pannovy’ yields were recorded from calendar week 17 to 40; data from weeks 17-32 were used for training, validation and testing the neural network, and weeks 33-40 were used to test the ability of the model as a yield predictor.
A layered digital neural network (LDDN), a type of dynamic neural network, was applied, which includes delay lines between the layers, so the output also depends on previous inputs and/or previous states of the network.
The best LDDN model was chosen on the basis of a suitable architecture (i.e., a minimum number of hidden layers and neurons connected to an output layer). Also, two statistical measures were used to examine the goodness of fit between measured and predicted yields, the correlation coefficient (R) and the mean absolute error (MAE). Results show that an LDDN with one hidden layer with four nodes, six delays for the inputs and four delays for the output variable gives the best performance.
The most influencing variable for tomato yield prediction was CO2 fixation, and the less important variable was transpiration.
This work shows the power of the LDDN for prediction purposes, which can be used to predict yields in both greenhouses one week in advance.
Authors
R. Salazar, D. Dannehl, U. Schmidt, I. López, A. Rojano
Keywords
transpiration, CO2 fixation, solar radiation, LDDN
Online Articles (33)