Application of Random Forest and Back Propagation Neural Network in Estimating Radiation-Based Reference Evapotranspiration Model in Gansu Province

Reference evapotranspiration (ETo) is an important factor in the water-saving agriculture and in the soil-plant-atmosphere continuum. Many machine learning methods have been introduced into predicting ETo. In order to improve the accuracy of radiation-based ETo models, this paper presents an ETo model called RFR based on random forest (RF). By taking the results of FAO Penman-Monteith (FAOPM) model as the standard, 4 radiation-based ETo models, including RFR model, Priestley-Taylor (PT) model, Makkink model and BPR model (radiation-based ETo model based on back propagation neural network) were evaluated in Yumen, Gansu Province. Results show that RFR model can fit the nonlinear mapping relationship between the radiation parameters and ETo well. The accuracy of RFR model (with RMSE of 0.81 mm/day) is superior to BPR model (with RMSE of 0.89 mm/day) and other compared empirical models. At the same time, training time of RFR model is shorter than BPR model. Therefore, the method for constructing ETo model based on random forest can effectively improve the accuracy of radiation-based ETo model in Gansu Province.