Estimation Model Of Reference Evapotranspiration In The Cultivation Facility Of Panax Notoginseng Based On Machine Learning

As a unique medicinal material in China,Panax notoginseng has brought good economic benefits to Yunnan Province.Panax notoginseng is sensitive to irrigation amount,too much or too little water is not conducive to its growth and quality.Therefore,as an important index for making irrigation decisions of Panax notoginseng,reference evapotranspiration（ET₀）becomes the key to explore water requirement of Panax notoginseng.In this study,meteorological data collected by instruments installed in the cultivation facility of Panax notoginseng in Luxi County,Yunnan Province,and downloaded from the local national weather station were used to analyze the main meteorological factors affecting ET₀ inside and outside cultivation facility of Panax notoginseng based on path analysis method,and explore the feasibility of estimating ET₀ in facility using meteorological factors outside the facility;Markov chain Monte Carlo method was used to calibrate the parameters of six empirical models for estimating ET₀,and the estimation performance of ET₀ in facility was evaluated;The parameters of three machine learning methods（support vector regression machine（SVR）,random forest（RF）and extreme learning machine（ELM））were optimized by Bayesian optimization（BO）algorithm,and three kinds of ET₀estimation models（BO-SVR,BO-RF and BO-ELM）in cultivation facility of Panax notoginseng were established.The main conclusions are as follows:（1）The response characteristics of ET₀ of Panax notoginseng field to meteorological factors in the shading facility were proved.In cultivation facility of Panax notoginseng,average relative humidity RH had the greatest influence on ET₀,showing a significant negative correlation（r=-0.959）,and the determining coefficients of other meteorological factors on ET₀from high to low were as follows:average wind speed u>maximum temperature T_max>net solar radiation R_n>minimum temperature T_min>mean temperature T_mean.ET₀in the facility was less affected by the net solar radiation and was mainly determined by the aerodynamic term.There was correlation between meteorological factors inside and outside the facility,which made the change trend of ET₀inside and outside the facility consistent on the whole,and the fitting effect of linear equation was well,with R²of 0.805.Therefore,it was feasible to estimate ET₀in the cultivation facility of Panax notoginseng by using meteorological factors outside the facility.（2）The adaptability of the improved simple model to estimate ET₀of Panax notoginseng field in shading facility was clarified.Outside the facility,sunshine duration n-out,average relative humidity RH-out and maximum temperature T_max-out had a great influence on ET₀（|r|>0.794）.Before calibrating parameters,Valiantzas model based on sunshine duration n-out,average relative humidity RH-out and maximum temperature T_max-out had the highest estimation accuracy of ET₀ outside the facility,with R²of 0.957.The parameters calibrated based on Markov chain Monte Carlo method deviate from the initial value slightly,and the parameters of some models have seasonal differences.After parameter calibration,Irmak model had the highest accuracy in estimating ET₀ outside the facility.In the seedling stage,flowering stage and fruiting stage of Panax notoginseng in 2019（calibration year）,R²was 0.932,0.979 and 0.922,respectively.It also had a good effect in estimating ET₀in the facility through the relationship between inside and outside the facility,with R²of 0.728,which was 13.93%higher than before the parameter calibration.Valiantzas model had the highest accuracy in estimating ET₀in the facility,with R²of 0.779.（3）The ET₀estimation model of Panax notoginseng field based on machine learning method was established.Among the six meteorological factors outside the facility,average relative humidity RH-out had the greatest effect on ET₀in the cultivation facility of Panax notoginseng,with a correlation coefficient of-0.935,while minimum temperature T_min-out had the least effect on ET₀,with a correlation coefficient of-0.264.The overall test accuracy of machine learning model was BO-ELM>BO-SVR>BO-RF,and BO-ELM model based on the combination of meteorological factors including average relative humidity RH-out,maximum temperature T_max-out,average wind speed u-out and average temperature T_mean-out had the highest estimation accuracy in estimating ET₀in the facility,R²was 0.926,RMSE was 0.071 mm/d and MAE was 0.055 mm/d,which was suitable for estimating ET₀in the facility with a few meteorological factors(RH-out and T_max-out),R²was 0.893,RMSE was 0.085 mm/d and MAE was 0.068 mm/d.The average running time of BO-SVR and BO-ELM was 1.01 s and 2.20 s,respectively.The computational cost of BO-RF was 28.87 s on average.Considering the calculation accuracy and cost comprehensively,the BO-ELM model could be used as the estimation method of ET₀in cultivation facilities of Panax notoginseng in the absence of meteorological factors.