Modeling daily reference evapotranspiration via a novel approach based on support vector regression coupled with whale optimization algorithm

• Published in: Agricultural water management Vol. 237; p. 106145
• Main Authors: Mohammadi, Babak, Mehdizadeh, Saeid
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.07.2020
• Subjects: algorithms; climate; evapotranspiration; Hybrid model; Iran; irrigation scheduling; Pre-processing; principal component analysis; Reference evapotranspiration; regression analysis; Support vector regression; Whale optimization algorithm; Iran; Reference evapotranspiration; Support vector regression; Hybrid model; Whale optimization algorithm; Pre-processing
• ISSN: 0378-3774; 1873-2283
• DOI: 10.1016/j.agwat.2020.106145

•Daily reference evapotranspiration was modeled at three stations in Iran.•Model inputs were: air temperatures, relative humidity, solar radiation, sunshine duration, and wind speed.•Pre-processing was used for optimal input identification.•Proposed an approach that couples support vector regression with whale optimization algorithm.•The support vector regression-whale optimization algorithm was better than the single support vector regression. In achieving water resource management goals such as irrigation scheduling, an accurate estimate of reference evapotranspiration (ET0) is critical. Support vector regression (SVR) was applied to the modeling of daily ET0 at three meteorological stations in Iran subject to different climates: Isfahan (arid), Urmia (semi-arid), and Yazd (hyper-arid). Different pre-processing approaches [relief (RL), random forests (RF), principal component analysis (PCA), and Pearson's correlation (COR)] served to determine the SVR’s optimal input combinations. While these approaches introduced different inputs to the SVR models, those drawn upon by the RF approach (i.e., RF-SVR) generated better results than other approaches. Models performance was evaluated using the root mean square error (RMSE), normalized RMSE (NRMSE), mean absolute error (MAE), coefficient of determination (R2), and the Nash-Sutcliffe efficiency (E). A novel hybrid model, coupling SVR with a whale optimization algorithm (WOA), was also developed and applied to daily ET0 modeling. The hybrid models outperformed the SVR-only models, with the hybrid RF-SVR-WOA model having the best performance.