Hybrid Neural Network Based Models for Evapotranspiration Prediction Over Limited Weather Parameters

• Published in: IEEE access Vol. 11; pp. 963 - 976
• Main Authors: Vaz, Pedro J., Schutz, Gabriela, Guerrero, Carlos, Cardoso, Pedro J. S.
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Agriculture; Algorithms; Artificial neural networks; Atmospheric models; Climate models; Computational modeling; Crops; Evapotranspiration; Hydrology; Irrigation; Meteorological parameters; Meteorology; Moisture content; Neural networks; Predictive models; public garden; Radiation; Recurrent neural networks; smart irrigation; Solar radiation; Water management; Weather forecasting; Weather stations
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2022.3233301

Evapotranspiration can be used to estimate the amount of water required by agriculture projects and green spaces, playing a key role in water management policies that combat the hydrological drought, which assumes a structural character in many countries. In this context, this work presents a study on reference evapotranspiration (<inline-formula> <tex-math notation="LaTeX">ET_{o} </tex-math></inline-formula>) estimation models, having as input a limited set of meteorological parameters, namely: temperature, humidity, and wind. Since solar radiation (SR) is an important parameter in the determination of <inline-formula> <tex-math notation="LaTeX">ET_{o} </tex-math></inline-formula>, SR estimation models are also developed. These <inline-formula> <tex-math notation="LaTeX">ET_{o} </tex-math></inline-formula> and SR estimation models compare the use of Artificial Neural Networks (ANN), Long Short Term Memory (LSTM), Gated Recurrent Unit (GRU), Recurrent Neural Network (RNN), and hybrid neural network models such as LSTM-ANN, RNN-ANN, and GRU-ANN. Two main approaches were taken for <inline-formula> <tex-math notation="LaTeX">ET_{o} </tex-math></inline-formula> estimation: (i) directly use those algorithms to estimate <inline-formula> <tex-math notation="LaTeX">ET_{o} </tex-math></inline-formula>, and (ii) estimate solar radiation first and then use that estimation together with other meteorological parameters in a method that predicts <inline-formula> <tex-math notation="LaTeX">ET_{o} </tex-math></inline-formula>. For the latter case, two variants were implemented: the use of the estimated solar radiation as (ii.1) a feature of the neural network regressors, and (ii.2) the use of the Penman-Monteith method (a.k.a. FAO-56PM method, adopted by the United Nations Food and Agriculture Organization) to compute <inline-formula> <tex-math notation="LaTeX">ET_{o} </tex-math></inline-formula>, which has solar radiation as one of the input parameters. Using experimental data collected from a weather station (WS) located in Vale do Lobo (Portugal), the later approach achieved the best result with a coefficient of determination <inline-formula> <tex-math notation="LaTeX">(R^{2}) </tex-math></inline-formula> of 0.977. The developed model was then applied to data from eleven stations located in Colorado (USA), with very distinct climatic conditions, showing similar results to the ones for which the models were initially designed (<inline-formula> <tex-math notation="LaTeX">R^{2}>0.95 </tex-math></inline-formula>), proving a good generalization. As a final notice, the reduced-set features were carefully selected so that they are compatible with free online weather forecast services.