关键词: 高标准农田, 参考作物蒸散, 经验模型, 机器学习模型, 蒸散模拟

Abstract:

Accurate simulation of reference crop evapotranspiration (ET₀) can provide scientific guidance for agricultural water resources management of high−standard farmland. However, issues such as data quality of microclimate stations pose some challenges for ET₀ estimation. To solve above problems, this study selected the daily meteorological observation data of 16 farmland microclimate stations with complete data records and 13 nearby national meteorological stations in Henan province from 2020 to 2023, adopted 13 typical empirical models and 8 machine learning models to estimate ET₀, and took Penman−Monteith (PM) model as the benchmark. The accuracy of each model was evaluated, and 15 types of ET₀ estimation combination schemes based on the optimal model were given to find accurate, suitable and simple alternative models to estimate ET₀. The results showed that in addition to wind speed (WS), the fitting degrees of average temperature (T_ave), maximum temperature (T_max), minimum temperature (T_min), mean air relative humidity (RH), vapor pressure deficit (VPD) and net solar radiation (R_n) observed between microclimate stations and national stations were generally higher than 0.654 (P<0.05). The correlation between ET₀ calculated based on the above two datasets was also higher, with R² of 0.880 and RMSE of 0.588mm. Among the 13 empirical models, the Valiantzas3 (Val3) model considering temperature, radiation, relative humidity and wind speed exhibited the best effect (R²=0.933, RMSE=0.461mm), followed by Jensen−Haise (JH) model considering radiation and temperature factors (R²=0.916, RMSE=0.774 mm). The overall accuracy of the temperature−based Hargreaves−Samani (HS) model was high (R²=0.817, RMSE=0.713mm), while the simulation accuracy of the mass−transfer based Penman (Pen), WMO and Trabert (Tra) models was lower and there were not recommended as the choice of ET₀ simulation. Among the 8 machine learning models, the simulation accuracy of Multilayer Perceptron (MLP) model was the best (R²=0.998, RMSE=0.059mm), and the importance order of each input parameter was: R_n>VPD>WS>T_max>T_ave>T_min>RH. Among 15 different model input parameter combination schemes based on the MLP model, the simulation accuracy of the machine learning model was generally better than that of the empirical model under the same input parameter conditions, with the combined model of R_n+T_ave+RH+WS being the best.

Key words: High?standard farmland, Reference crop evapotranspiration, Empirical model, Machine learning model, Evapotranspiration simulation