Comparison of the Efficiency of Calibrated Combined Relations and Intelligent Neural Systems in Estimating Evaporation from Free Water Zones

Background and Objectives
Evaporation is one of the most important factors in the hydrological cycle and is one of the determinants of energy equations at the ground level and water balance, which is estimated in various fields such as meteorology, hydrology, agriculture, and water resources management. Evaporation is also one of the main causes of water loss and stress on water resources. Therefore, knowing its amount as one of the hydrological variables is very important in agricultural research and soil and water conservation and modeling. Evaporation is a physical process that has a direct and close relation with atmospheric factors, the most important of which are temperature, wind speed, relative humidity and solar radiation. Researchers have been able to analyze evaporation using mathematical and empirical methods and their combination, as well as using intelligent neural methods. Due to the importance of evaporation in the water cycle and its effect on the quantity and quality of surface water resources, the study and accurate knowledge of this phenomenon is one of the important issues in the study of water resources. Using pan evaporation is one of the most common methods of estimating evaporation. But in most areas, the number of evaporating stations is not enough and they do not have suitable spatial distribution. Therefore, indirect methods such as hybrid relations, intelligent neural systems, data mining methods and remote sensing techniques have been considered by researchers.
Methodology
In the present study, the evaporation of free water zones in the Urmia Lake basin has been estimated. For this purpose, the efficiency of combined empirical methods including deBruin, Tichomirov, Penman and Meyer as well as intelligent neural methods including artificial neural networks (ANN), random forests (RF) and gradient boosted trees (GBT) were compared and evaluated using statistical indices of R, NRMSE, MAPE and also Taylor diagram. Moreover, in order to increase the accuracy and efficiency of the combined methods, these relations were calibrated for the Urmia Lake basin. In order to evaluate the different combinations of meteorological variables to estimate the evaporation of free water zones in intelligent neural systems, 14 scenarios were considered with the aim of increasing the accuracy of evaporation estimation. In these scenarios, various combinations of meteorological parameters were defined that were used as variables of the combined empirical relations to estimate evaporation of free water zones. Also, pan evaporation data were used to estimate the rate of evaporation of free water zones by applying the pan coefficient and the obtained results were used as a basis for evaluating combined methods and intelligent neural systems.
Findings
The results showed that among the studied combined methods at six considered stations, the deBruin method is more accurate than other methods. Only in Tekab station, the Meyer method with NMRSE value of 30.00% and MAPE of 19.99% had higher accuracy. After calibrating the relations, the deBruin method also had the highest accuracy in all stations compared to other relations. Among the intelligent neural methods in 4 of 7 studied stations, the ANN method was introduced as the best and most accurate intelligent method for estimating evaporation of free water levels. In Maragheh, Mahabad and Sarab stations, RF method had the highest accuracy, while in all of the stations, the GBT method had the weakest performance.
Conclusion
Despite the overall improvement in the results of the evaporation estimation and the reduction of the error values of the calibrated empirical combined relations, the NMRSE values indicated different efficiencies of the combined relations in estimating the evaporation of free water zones. So, the calibrated combined relations were not accurate at any of the stations. Moreover, evaluating the results of intelligent neural methods indicated the high accuracy of them compared to combined relations in estimating evaporation of free zones of water. Also, the obtained results showed that the temperature and radiation parameters in the model obtained from the best scenario of intelligent methods have been used in all stations, which indicated the importance of these two parameters in evaporation modeling. Also, the results showed that although the calibration of the relations generally improved the accuracy of the combined relations; however, according to the statistical analysis, the combined relations did not have the suitable accuracy in estimating evaporation. Therefore, the use of intelligent neural systems in estimating evaporation of free zones of water was recommended. Among all of the studied methods, the ANN method had the highest accuracy in estimating the pan evaporation. Thus, this method was introduced as an accurate model in 4 stations with NRMSE values less than 10%.