| Under the background of global climate change,drought has become the most severe climate disaster of restricting regional cropland production,which is easy to reduce crop production.Evapotranspiration is an important indicator of monitoring drought,an important factor of hydrological cycle and water resource management,and a key factor of matter and energy exchange between soil-crop-atmosphere system.It has a very close relationship with crop physiological activities and yield.Accurately estimating evapotranspiration can provide useful information for water management and sustainable agricultural development,which is of great significance to take effective drought resistance and disaster reduction measures,and is of great practical significance to improve the yield of cropland crop.The purpose of this paper is to develop the hybrid model that can accurately obtain global or regional evapotranspiration,the main works include:（1）Six different hybrid ET models were constructed.The six hybrid models were based on six traditional ML algorithms,namely K-nearest Neighbor algorithm（KNN）,Random Forest（RF）,Support Vector Machine（SVM）,e Xtreme Gradient Boosting algorithm（XGBoost）,Artificial Neural Network（ANN）,and Long Short-term Memory（LSTM）.The models were constructed using observed data of 17 eddy covariance flux sites of cropland over the globe.The models parameters were further optimized after training the models to improve the performance of the models.（2）Each hybrid model was assessed to estimate cropland ET with ten different input factors combinations.Meteorological data and one,two,three,and four remote sensing data were input into the machine learning method to build a stomatal conductance model,and then Penman-Monteith（PM）equation was used to estimate evapotranspiration.（3）The accuracy of six machine learning-based hybrid models for estimating evapotranspiration with using the same input variables was evaluated in order to obtain a better method for estimating evapotranspiration.（4）Medlyn-Penman-Monteith model was constructed.The improved hybrid machine learning model was compared with the Medlyn-PM model to obtain a model that is suitable for different environmental conditions,improves the accuracy of regional scale evapotranspiration simulation,and provides method support for accurate estimation of regional scale evapotranspiration.（5）The accuracy of ANN-PM model under dry climates was studied.The correlations coefficient（r）between simulation and observation values of the ANN-PM model was used to evaluate the model performance.We studied whether the model could capture the time-series changes of evapotranspiration at the dry sites by comparing the time-series diagrams of observed and simulated values by the ANN-PM model.The main results are as follows:（1）The performance of the hybrid models of using four remote sensing factors was best.Compared with the hybrid models with using three or four remote sensing factors,the performance of the hybrid models using two remote sensing factors is similar.（2）The predicted evapotranspiration based on ANN hybrid model was consistent with the observed value,showing the best accuracy(RMSE=18.67-20.69 W m-2,r=0.90-0.94).（3）The performance of the hybrid machine learning model（ANN-PM model）was superior to Medlyn-PM model in estimating evapotranspiration(RMSE=19.23-19.71 W m-2,r=0.93).（4）The performances of the ANN-PM model at the dry sites were reasonable（the average r was 0.87）.The ANN-PM model can capture the time-series changes of ET at the dry sites well. |
