Estimation Of Reference Evapotranspiration With Machine Learning Methods Under Future Climate Change

Reference Evapotranspiration（ET_o）is an important part of agricultural water cycle and balance.Accurate estimation of reference evapotranspiration（ET_o）is essential for agricultural water resource management,crop water demand estimation,and optimal irrigation scheduling.The FAO-56 Penman-Monteith model,which is recommended by the Food and Agriculture Organization of the United Nations（FAO）,is regarded as the standard model to calculate ET_oand can obtain the highest estimation accuracy.However,its application is usually limited by the lack of required complete meteorological dataset in many regions.In order to explore the ET_o estimation methods that can replace the FAO-56 Penman-Monteith model under the condition of limited meteorological data,this study used the daily meteorological data of 63meteorological stations with radiation observations in four different climate zones of China during 1991-2017.The ET_o values calculated by the FAO-56 Penman-Monteith model was regarded as the standard values in this period.The ET_o estimation accuracies of seven machine learning models were compared under seven different combinations of input meteorological factors.Then,three machine learning models were developed with only three input meteorological factors at 696 meteorological stations across entire China mainland,whose ET_o estimation accuracies were compared with the corresponding ET_o empirical models.Finally,based on the meteorological data of 27 GCMs（Global climate models）of CMIP6（the Coupled Model Intercomparison Project Phase 6）and two emission scenarios（SSP245 and SSP585）,we analyzed the changes of meteorological factors at 696 meteorological stations in China mainland,as well as the spatial distribution characteristics of ET_o in the future.The main conclusions of this study were drawn as follows.（1）The meteorological factors of R_s,T_max and R_a had strong correlations with ET_o.With the reduction of meteorological input factors,the accuracy of ET_o estimation by the machine learning models decreased.The best ET_o estimation accuracy was obtained when the combination of input meteorological facors was T_max+T_min+U₂+RH+R_s,followed by the combination of T_max+T_min+U₂+RH+R_a.The combination of T_max+T_min+R_a,which only had three input variables,had the lowest estimation accuracy.（2）When estimating ET_o with seven different machine learning models（i.e.,RF,SVM,XGBoost,Light GBM,MLP,MLR,and SG.MLR）,the estimation accuracies of machine learning models were much higher than that of the multiple linear regression method.Especially,the SG.MLR,SVM and MLP models showed better performance.The average R²value of the SG.MLR model was 0.28%-0.87%higher than those of the single machine-learning models,whereas the values of RMSE,MAE,and NRMSE were reduced by 3.29%-12.81%,2.40%-13.57%and 3.28%-12.75%,respectively.This suggests that the stacking ensemble method had greater potential to improve the prediction performance of single machine learning model.（3）The ET_o estimation accuracies of machine learning models were higher than those of the corresponding traditional empirical models.Among the empirical models,only the Lobit model and the Abtew model had better estimation performance for ET_o estimation in some areas.In the four climate zones across entire China mainland,the ET_o estimation accuracies of Subtropical Monsoon Zone（SMZ）,Mountain Plateau Zone（MPZ）and Temperate Monsoon Zone（TMZ）were generally higher than that of the Temperate Continental Zone（TCZ）.The average annual ET_o values estimated by the machine learning models were very close to those calculated by the Penman-Monteith model,while empirical models typically overestimated ET_o values for most sites.In the absence of complete meteorological data,it was recommended to use machine learning models to estimate daily ET_o in different regions of China mainland.（4）Under future climate scenariso,both the maximum temperature T_max and minimum temperature T_min showed increasing trends,and the increases under the SSP585 scenario were greater than those under the SSP245 scenario.As time went on,the increasing magnitude of T_max and T_min also increased.Solar radiation R_s showed a downward trend in the MPZ and TCZ regions,and an upward trend in the SMZ and TMZ regions.Rainfall generally showed an increasing trend,but the level of increase was significantly different in each region,with obvious characteristics of higher in the east and lower in the west.（5）In the future,the ET_o of the four climatic zones would increase remarkably in China mainland,gradually increasing from 2040s to 2090s and reaching the maximum in 2090s under the SSP585 scenario.Especially,ET_o would increase by 200 to 400 mm year^-1 in the SMZ and TCZ regions.The increase of ET_o was greater than the increase of rainfall in the MPZ and TCZ regions,while the increase of ET_o was equal to or slightly larger than the increase of rainfall in the SMZ and TMZ regions.Thus,dry regions would become drier while wet regions would become wetter across China in the future.