Document Type : Research Paper
Materials and Methods
In the current research, the meteorological data of ten years (2009–2019) from the synoptic stations of Zanjan, Mahneshan, Khodabande and Khoramdare were used. The input of multiple linear regression (MLR), gene expression programming (GEP) and random forest (RF) models includes four meteorological parameters (average temperature, average relative humidity, wind speed and net radiation). In order to evaluate the accuracy of the models, the daily reference evapotranspiration of FAO Penman Monteith (FPM) was chosen as the standard method. To evaluate the performance of the models, one-third and two-thirds of the data were used for training and validation, respectively. Also, to compare the performance of the machine learning methods with each other, the evaluation criteria including the Root Mean Square Error (RMSE), Coefficient of determination (R2), Scatter Index (SI) and Wilmot Index (WI) were calculated and time series, dispersion and violin charts were drawn. On the other hand, the geostatistical technique was used for estimating the reference evapotranspiration in unmeasured points for zoning because of the limited number of available synoptic stations in Zanjan province. In order to analyze the spatial changes of reference evapotranspiration, its zoning was carried out after the highest-performing machine learning models were chosen, considering that it has not been evaluated in this manner before in the study area.
Findings
The random forest (RF) model in Zanjan station has good performance with WI = 0.991, SI = 0.109, and RMSE = 0.373(mm/day), R2= 0.97; the gene expression programming (GEP) model in Khodabande station has the best performance with WI = 0.966, SI = 0.124, RMSE = 0.478(mm/day), and R2 = 0.967 and the Multiple linear regression (MLR) model has the best performance in Zanjan station with WI = 0.955, SI = 0.08, RMSE = 0.276(mm/day), and R2 = 0.983. After the random forest model, the gene expression programming model in Khodabandeh and Mahenshan stations and the multiple linear regression model in Khoramdare stations ranked second. In all stations, except for Zanjan station, where the MLR model performed well, the RF model has high accuracy and less error in estimating daily reference evapotranspiration. Given that the RF model has high accuracy in predicting daily reference evapotranspiration, the values of this model were used for evapotranspiration zoning using the inverse distance weighting (IDW) method. The created map showed that Reference evapotranspiration in Zanjan province is in the range of (3.466-4.273) mm/day. In the northwestern region of this province, its amount increases, and its highest amount is in the range of (4.020-4.273) mm/day at the Mahneshan station, which shows that the amount of precipitation and climatic conditions affect the amount of evaporation-transpiration.
Conclusion
Taking into consideration the important role of accurate evapotranspiration prediction in the design, planning, and management of irrigation networks and systems, in the present research, the synoptic data of 4 stations in Zanjan province for 10 years, including average temperature, average relative humidity, wind speed and net radiation were input to the model. Multiple linear regression (MLR), gene expression programming (GEP) and random forest (RF) were used to estimate daily reference evapotranspiration. According to the values of statistical indicators, the results showed that all three models have high accuracy and less error in estimating daily reference evapotranspiration, but the random forest (RF) model was chosen as the best model with a small difference. Based on the distribution of the values predicted by the machine learning models and comparing it with ET0-FPM values, the random forest (RF) model has less difference. Also, the closest violin chart to the ET0-FPM data is the random forest model chart, which shows the superiority of the random forest model over the other models. The results of zoning using the inverse distance weighting (IDW) method of this model showed that the rate of evapotranspiration is higher in the northwest of Zanjan province, especially in Mahenshan Station with (4.020-4.273) mm/day.
Acknowledgement
We acknowledge that the meteorological dataset utilized in this investigation was provided by the Islamic Republic of Iran Meteorological Organization (IRIMO).
