Comparing of Generalized Linear Models, Random Forest and Gradient Boosting Trees in Estimation of Reference Crop Evapotranspiration (Case Study: The Sistan Plain)

The estimation of reference crop evapotranspiration (ETo) is one the important factors in hydrological studies, irrigation planning, and water resources management. This study attempts to explore the possibility of predicting this key component using three different methods in the Sistan plain: Generalized Linear Models (GLM), Random Forest (RF) and Gradient Boosting Trees (GBT). The maximum and minimum temperature, mean temperature, maximum and minimum humidity, mean humidity, rainfall, sunshine hours, wind speed, and pan evaporation data were applied for years between 2009 to 2018. Using various networks, the ETo as output parameter was estimated for different scenarios including the combination of daily scale meteorological parameters. In order to evaluate the capabilities of different models, results were compared with the ETo calculated by FAO Penman-Monteith as the standard method. Among studied scenarios, M1 covering the maximum number of input parameters (10 parameters) showed the highest accuracy for GBT model, with the lowest RMSE (0.633) and MAE (0.451) and the maximum coefficient of regression (R = 0.993). Air temperature was found as the most sensitive parameters during sensitivity analysis of studied models. It indicated that accuracy and precision of temperature data can improve the results. Application of the GBT model could decrease the time consumed to run the model by 70%. Therefore, the GBT model is recommended for estimation of ETo in the Sistan plain.