Abstract:
The reference crop evapotranspiration (ET0) is a kernel parameter for precise irrigation management of crops and agriculture efficient water use. In order to improve the accuracy of the ET0 prediction in the hilly area of central Sichuan with missing meteorological data in some area, 15 different prediction models based on multivariate adaptive regression splines (MARS) were established by using different meteorological factors. The daily meteorological data of 11 representative meteorological stations from 1961 to 2016 were analyzed by the MARS models. These data were divided into training set and test set in a ratio of 7:3, and the simulation results of the MARS models were statistically evaluated using the calculation results of the FAO 56 Penman-Monteith model as a standard. In the statistical evaluation, 4 statistical parameters were obtained by the prediction sequence and the calculation result of the FAO 56 P-M model. They were root mean square error (RMSE), mean relative error (MRE), Nash efficiency coefficient (NSE), and R2. The value of the index above were used to calculate a score for evaluating the prediction accuracy of the models, and rank the models based on the scores. Then the results were compared with other ET0 prediction models and the applicability of the models in the hilly area of central Sichuan was evaluated by the portability analysis. The results showed that the full MARS model with 6 input parameters had the highest accuracy. Decreasing 1 input of relative humidity, the model still had the higher accuracy, ranking No 1 based on comprehensive performance indicator (CPI), which was same with the full model ranking. Reducing continually 1 input of sunshine duration still yielded the high simulation accuracy with NSE and R2 higher than 0.985. Further decreasing 1-2 input, the model NSE and R2 still were higher than 0.9. Among these models, the model with 2 inputs of radiation and wind speed was the most easy to use since the radiation could be calculated and only wind speed was required to measure. Radiation and wind speed were the keys to determine the regional adaptability of machine learning models. Radiation contained the geographic and temporal information of the site, which made it a key factor in the MARS models to distinguish the differences in geographical environment. On the other hand, radiation could compensate for the negative impact caused by the lack of sunshine duration on the prediction accuracy of the MARS models. The wind speed was more important than the other meteorological factors because the response of MARS models were more sensitive to it. Compared with the Irmak-Allen Model, the Irmak Model, and the Hargreaves-M4 Model, the MARS6, MARS7, and MARS8 improve the accuracy. Under the same meteorological factors input, the MARS models had a stronger simulation ability for ET0 than the existing empirical models; Through the portability analysis, the MARS model with 4 input parameters of radiation, maximum and minimum temperature and wind speed maintained high precision with NSE and R2 both higher than 0.985, RMSE 0.121-0.193 mm/d and MRE 2.7%-4.1%. In sum, the MARS model realized the deletion and replacement of meteorological factors, reduced dependence of ET0 forecasting on meteorological data, and maintained a relatively high forecasting accuracy and wide applicability. The MARS was recommended as a reliable ET0 prediction model in the hilly area of central Sichuan.