Research And Application Of Global Evapotranspiration Simulation Algorithm Driven By Remote Sensing Near Infrared Vegetation Inde

Evapotranspiration(ET)is the sum of vegetation transpiration and surface soil evaporation.With the gradual increase of climate warming,atmospheric circulation is affected,resulting in changes in the temporal and spatial pattern of ET.Accurately quantifying of global ET is a crucial approach to understand ecosystem water cycle and climate change.However,ET models vary in complexity and precision,with various limitations in regional expansion.The recently developed Near-Infrared Reflectance of Vegetation(NIRv)is potentially useful for improving the accuracy and applicability of ET models,as well as reducing complexity in ET modeling.In this study,based on the Python language,utilizing eddy flux site data and remote sensing data,three NIRv-driven ET models were constructed to estimate ET of 11 vegetation types at the global scale.The main innovations are as follows:(1)Two ET models,NIRv-Penman-Monteith(NIRv-PM)and NIRv-underlying Water Use Efficiency(NIRv-uWUE),were constructed.The models were trained and validated on 32 cropland sites,and the models were compared with six mature ET models on validation set and arid sites.The results show that both NIRv-PM and NIRv-uWUE were able to reasonably simulate cropland ET.The accuracy of NIRv-PM in estimating ET(R2=0.741,RMSE=5.638 mm/8d,RMSEs=1.519 mm/8d,RMSEu=5.43 mm/8d)is higher than that of the NIRv-uWUE model(R2=0.674,RMSE=6.275 mm/8d,RMSEs=2.091 mm/8d,RMSEu=5.916 mm/8d)and the error is smaller.The performance of the NIRv-PM model is better than that of the mature models,and it is still stable in arid climate.(2)The NIRv univariate linear regression method and light use efficiency model were used to estimate the total Gross Primary Productivity(GPP)of 11 vegetation types,and the consistency between estimated GPP and observed GPP was evaluated.The results show that NIRv and GPP are highly correlated within all vegetation types,and the simulation ability of the light use efficiency model is significantly better than the linear regression method.The average R2 of all vegetation types is 0.812 and 0.764,and the average RMSE is 0.978 g C m-2 d-1 and 1.031 g C m-2 d-1,respectively.(3)Two machine learning methods,Artificial Neural Network and Random Forest,were used to optimize the key parameter "PT coefficient" in NIRv-PM model,so as to improve the interpretation of the advection process by multiple effect factors.The calculated PT coefficient estimations were substituted into the ET model for validation.The results show that high simulation accuracy of PT coefficient can be achieved using machine learning models,and the accuracy of Random Forest is even higher.(4)The feasibility of estimating ET with the improved NIRv-PM model was validated using time series data from 115 flux sites at the global scale.The results show that the improved NIRv-PM model(the average value of R2,RMSE,RMSEs and RMSEu is 0.732,0.709 mm/d,0.387 mm/d,and 0.564 mm/d respectively)can accurately estimate the ET of the complex underlying surface at the global scale.Compared with the original NIRv-PM model(0.699,0.685 mm/d,0.463 mm/d and 0.489 mm/d,respectively),the interpretation ability of ET changes has been effectively enhanced.Afterwards,the reliability of the improved NIRv-PM model in North China Plain(NCP)were validated using remote sensing raster data,and the temporal and spatial distribution characteristics of ET in NCP were analyzed.The above demonstrate that the improved NIRv-PM model has wide applicability on global scale,providing a simple and robust method for estimating global ET.