Application of Tree and Kernel- Based Models for Estimating Daily Reference Evapotranspiration in Humid and Arid Regions of Iran

The gradual increase in the world’s population requires continues increase in agricultural production. Climate change is one of the challenges of our society and frequent droughts affect large areas of the world, which requires more accurate management of water resources, both globally and in local catchments. Accurate estimation of components of the hydrological cycle is essential for proper irrigation scheduling. Most of the precipitation received by the earth is returned to the earth’s atmosphere by the process of evapotranspiration. On the other hand, because every process that takes place in the plant is dependent on water and one of the most common uses of water in the plant is evapotranspiration, so reducing amount of the water will have adverse effects on photosynthesis, crop production, product quality, etc. The complex and nonlinear relationship between the factors affecting the process of evapotranspiration, has caused researchers today to use new methods to accurately identify and predict this parameter. Reference evapotranspiration is a concept that uses the crop coefficient to obtain the actual water requirement. According to the FAO proposal, the FAO- Penman- Monteith equation was introduced as a benchmark method for calculating reference evapotranspiration values when measurements of this parameter are not available and there is no access to lysimetric data. One of the major advantages of this model is its physical basis and global validity, but this equation needs a large number of meteorological parameters that are often not available, instead empirical equations with low meteorological variables or modern methods such as artificial intelligence and machine learning methods can be used.

In this study, meteorological data related to two stations of Astara located in the humid region and Sirjan located in the arid region of Iran in the period of 2000-2020 were studied to predict the crop evapotranspiration values. As mentioned, the FAO- Penman- Monteith method has used as a standard method for calibration and evaluation of the other functional equations and machine learning methods. In this study, four types of empirical equations including Hargreaves –Samani, Makkink, Turk and Dalton were evaluated against the FAO- Penman- Monteith model. Also, modelling was performed using Support Vector Regression, Random forest and M5P Tree model. In this study, 70% of data were considered for training and 30% for testing. Finally, statistical parameters including root mean squared error (RMSE), correlation coefficient (R), scatter index (SI), Nash-Sutcliffe coefficient (NS) and Wilmot index (WI) were used to determine the performance of each mentioned methods in estimating reference evapotranspiration values.

Using different meteorological parameters in accurate prediction of evapotranspiration using 4 combined scenarios, calibration calculations were performed on 70% of data and validation calculations were performed on 30% of testing data implementing Weka software. The obtained results showed that the SVR3 and M5P3 models in Astara station with all meteorological parameters and having R= 0.993, RMSE= 0.201 and also, the SVR3 model in Sirjan station with R= 0.982, RMSE= 0.410 compared to the studied empirical methods provided better results in estimating the reference evapotranspiration and scenario 3 with all meteorological parameters was introduced as the top scenario. Among the empirical methods, Hargreaves- Samani was superior to some models only in Astara station. At Sirjan station, none of the empirical models performed better than the machine methods.

Accurate estimation of reference evapotranspiration in water resource management is essential. In this study, meteorological data from Astara and Sirjan stations were used to evaluate the ability of machine learning methods including SVR, RF and M5P to estimate the values of reference evapotranspiration and compared the results with empirical methods. The results showed that the high accuracy of the SVR3 model in both stations and in the next position M5P3 model for humid area. Empirical methods except Hargreaves- Samani had poor performance compared to data- driven models. Finally, the use of SVR and M5P methods in irrigation scheduling is recommended.