Comparison Of Machine Learning And Empirical Models For Estimating Spring Maize Evapotranspiration In The Case Of Shenyang Areas

Accurate estimation and simulation of crop evapotranspiration and actual water consumption have great significance for management and planning of irrigation practices.Based on the meteorological data of Shenyang area from 1971 to 2014,four machine learning models,named artificial neural network(BP),random forest(RF),support vector machine(SVM)and ensemble learning based on Adaboost algorithm,are established to simulate and estimate reference crop evapotranspiration.At the same time,estimation accuracy of reference crop evapotranspiration using above machine learning models with different input features is compared and analyzed.In order to further explore the applicability of machine learning model for estimating the actual water consumption of maize under non-conventional planting conditions(drip irrigation under mulch),machine learning models for estimating the water consumption of maize under mulch drip irrigation was established based on the datasets of maize under mulch drip irrigation at Liaoning Irrigation Center Test Station.The main results are as follows:(1)The absolute errors of RF1,BP1,SVM1 and Ada Boost1 in Shenyang are 0.54?0.72,0.52?0.74,0.53?0.70 and 0.56?0.64,respectively,which are lower than empirical formulas.It shows that when temperature datasets are available,the four machine learning models can accurately estimate reference crop evapotranspiration.Among them,Adaboost1 has the highest estimation accuracy.(2)The R~2 of RF2,BP2,SVM2 and Adaboost2 models for estimating reference crop evapotranspiration in Shenyang area are all greater than 0.8,and the absolute errors are0.40?0.57,0.41?0.57,0.36?0.53 and 0.52?0.54,respectively.The results show that the accuracy of the model has been significantly improved by adding the solar shortwave radiation.Moreover,the support vector machine model is better than the ensemble learning model.(3)The absolute errors of RF3,BP3,SVM3 and Adaboost3 models are 0.26?0.32,0.08?0.14,0.06?0.07 and 0.06?0.22,respectively.In addition,when complete meteorological data are available,the R~2 and NSE of four machine learning models are between 0.95?1.0 and0.94?1.0,respectively.The results indicate that the above four machine learning models can accurately estimate the reference crop evapotranspiration using complete meteorological data in Shenyang area,and the SVM3 model has the highest estimation accuracy.(4)Comparing the simulation results of four machine learning models for estimating water consumption of maize under drip irrigation with plastic film mulching,Adaboost model had the highest estimation accuracy,followed by support vector machine and random forest model,and then BP neural network model.However,the estimation errors of the four machine learning models are within the allowable range,indicating that machine learning models can accurately simulate and estimate the water consumption of maize under unconventional conditions(drip irrigation under mulch).(5)The estimation results of crop coefficient show that among the four machine learning models,Adaboost and SVM model were performing best in estimating crop coefficient,indicating that the four machine learning models can replace FAO recommendation method to estimate maize crop coefficient under drip irrigation under mulch.The results of estimating water consumption of maize showed that the accuracy of machine learning models can be expressed as SVM>BP>Adaboost model>FAO>RF in the early stage,in the fast growing stage,Adaboost model>SVM>RF>BP>FAO;in the middle stage,Adaboost model>BP>SVM>RF>FAO;in the late stage,BP>Adaboost>FAO>SVM>RF.For total water consumption,Adaboost model>BP>SVM>FAO>RF.Generally,Adaboost model has the highest accuracy and FAO recommendation method is the worst.