Comparative analysis of kernel-based versus ANN and deep learning methods in monthly reference evapotranspiration estimation

Timely and accurate estimation of reference evapotranspiration (ET0) is indispensable for agricultural water management for efficient water use. This study aims to estimate the amount of ET0 with machine learning approaches by using minimum meteorological parameters in the Corum region, which has an arid and semi-arid climate and is regarded as an important agricultural centre of Turkey. In this context, monthly averages of meteorological variables, i.e. maximum and minimum temperature; sunshine duration; wind speed; and average, maximum, and minimum relative humidity, are used as inputs. Two different kernel-based methods, i.e. Gaussian process regression (GPR) and support vector regression (SVR), together with a Broyden–Fletcher–Goldfarb–Shanno artificial neural network (BFGS-ANN) and long short-term memory (LSTM) models were used to estimate ET0 amounts in 10 different combinations. The results showed that all four methods predicted ET0 amounts with acceptable accuracy and error levels. The BFGS-ANN model showed higher success (R2=0.9781) than the others. In kernel-based GPR and SVR methods, the Pearson VII function-based universal kernel was the most successful (R2=0.9771). Scenario 5, with temperatures including average temperature, maximum and minimum temperature, and sunshine duration as inputs, gave the best results. The second best scenario had only the sunshine duration as the input to the BFGS-ANN, which estimated ET0 having a correlation coefficient of 0.971 (Scenario 8). Conclusively, this study shows the better efficacy of the BFGS in ANNs for enhanced performance of the ANN model in ET0 estimation for drought-prone arid and semi-arid regions.