A Method to Estimate Evapotranspiration in Greenhouse Conditions by Artificial Neural Networks Using Limited Climate Parameters

Abstract

Abstract Precise estimation of evapotranspiration (ET) within greenhouse environments assumes pivotal significance in the context of effective agricultural water resource management. It has an important influence on rational irrigation management and water conservation. The present study estimates evapotranspiration by artificial neural networks (ANNs) using limited climate parameters with data from Oct.2016 to Nov.2017 from an experimental greenhouse. Using a sigmoid transfer function, two ANN models, 2-5-1 structure and 4-9-1 structure, were established through the algorithm of multi-layer feed-forward and back-propagation. At the same time, moisture sensors installed at different depths of substrate were also used to calculate the amount of evapotranspiration. The standard reference evapotranspiration was provided by a microlysimeter system, which used an electronic weighing scale to continuously sample the amount of water supplied and lost during the experiment. The ANN (2-5-1) model estimated ET with an RMSE of 0.0915 L/d and an R2 0.9201, with two neurons in the input layer corresponding to the daily mean temperature and daily mean humidity, five neurons in the hidden layer and one neuron in the output layer corresponding to the reference evapotranspiration. The ANN (4-9-1) model estimated ET with an RMSE of 0.0592 L/d and an R2 of 0.9622, with four input climate parameters: daily maximum temperature, daily minimum temperature, daily mean temperature, and daily mean humidity, but it had nine neurons for the hidden layer. The results of linear regression analyses of ET estimation between moisture sensors and actual measurement show that the accuracy of moisture sensors is less than the ANN models (RMSE, 0.1129 L/d; R2, 0.8749). Therefore, these results confirmed the ability of ET estimation by ANN models using limited climate parameters in greenhouse conditions.