Estimation Of Solar Radiation And Reference Crop Evapotranspiration Based On Machine Learning Methods

The determination of crop water requirements is the key to water saving in agriculture.Global solar radiation?R_s?and reference crop evapotranspiration?ET₀?are important influencing factors for crop water requirements.Due to the limitations of observation costs and other reasons,meteorological data are usually used to estimate R_s and ET₀ instead of direct observation.The most commonly used methods are Angstrom-Prescott?A-P?formula for R_sand Penman-Monteith?P-M?formula for ET₀ estimation.In this study,China mainland was divided into four different climatic regions,namely the mountain plateau zone?MPZ?,the subtropical monsoon zone?SMZ?,the temperate monsoon zone?TMZ?,and the temperate continental zone?TCZ?.Then,we used measured solar radiation data from 80 stations that have solar radiation observations and meteorological observation data from 839 common meteorological stations,to obtain the a and b coefficients of the A-P formula across the country,based on the machine learning algorithms and existing empirical formulas.The optimal combinations of input meteorological factors for solar radiation estimation through different methods in different regions were verified,and the effects of solar radiation accuracy on the estimation of ET₀ based on the P-M formula were also explored.The main results and conclusions of this research were listed as follows.?1?Correlation analysis showed that the coefficient a of the A-P formula was correlated with many meteorological elements,and it could be more accurately estimated,while the coefficient b was relatively difficult to be estimated.Extra-terrestrial radiation?R_a?had a strong correlation with the potential hours?N?of sunshine.Several variables related to temperature(maximum temperature,T_max;minimum temperature,T_min;average temperature,T_mean;and diurnal temperature range,?t)also had strong correlations.They could be mutually interchangeable in the process of R_s estimation to some degree.?2?In the existing A-P parameter estimation models,altitude?Z?,latitude???,percentage of sunshine?n/N?and temperature?T?could be used to effectively estimate the coefficients.In this study,the support vector machine?SVM?method based on elevation and latitude achieved good results in a and b coefficient estimation.The correlation between coefficient a and altitude was obvious,generally increasing from southeast to northwest.The distribution of coefficient b was messy,generally increasing from northeast to southwest.?3?In different climatic regions of China mainland,the optimal combination of input meteorological factors required to estimate R_s was also different.We suggested that the optimal combination of input meteorological factors for a specific region should be determined according to its local climate characteristics.Three meteorological factors of sunshine hours?n?,extraterrestrial radiation?R_a?,and air temperature?T?had greater impacts on the solar radiation estimation.Adding the factor of precipitation could obviously improve the estimation accuracy in humid regions,but not remarkably in arid regions.Wind speed had very little influence on solar radiation estimation.In general,compared with other methods,the machine learning methods showed their advantages of flexible combinations of input meteorological factors and high estimation accuracy.Daily global solar radiation could be more accurately estimated with the SVM or Extreme Learning Machine?ELM?methods,which were driven by the optimal combination determined in this study.?4?When the P-M formula was used to estimate daily ET₀ in whole China mainland,if there were no measured radiation data,the coefficients of the A-P formula needed to be selected reasonably.Use of appropriate A-P coefficients in humid areas in southern China could especially reduce the ET₀ estimation error.Among several commonly used empirical models for daily ET₀ estimation,the Hargreaves-Samani method was superior when there were only temperature data.The model based on the SVM algorithm had higher accuracy than the traditional empirical model in ET₀ estimation,which was worthy of promotion in the future.