Regional Winter Wheat Yield Estimation Based On Bayesian Posterior Ensemble Data Assimilation

As two critical scientific methods for regional yield estimation and prediction,remote sensing methods and crop growth models have been widely used and verified.Data assimilation technology combines the advantages of remote sensing large-area information acquisition capabilities with crop growth models’mechanism and process characteristics.Thus,it is one of the essential methods to improve the accuracy of regional yield estimation.However,the remote sensing assimilation of the current crop model still has the problems of unclear quantification of uncertainty and low assimilation efficiency of high spatial resolution data.Compared with conventional parameter optimization methods such as trial and error methods and least squares methods,Markov Chain Monte Carlo(MCMC)method based on the Bayesian theory has significant advantages in calibrating complex crop growth models and quantifying parameter uncertainty based on a small amount of ground observation data.Regarding the above problems and focus on regional winter wheat yield estimation,in this dissertation,an Ensemble Kalman Filter(EnKF)assimilation estimation method that introduces a Bayesian posterior parameter set and a fast assimilation method based on a Bayesian rear prediction set suitable for 10-m resolution remote sensing data are constructed,establishing a technical process for regional winter wheat yield estimation at 10-km,250-m,and 10-m spatial resolution under the Bayesian probability framework.This dissertation focuses on improving the accuracy and efficiency of data assimilation-based yield estimation by combining the model uncertainty of MCMC calibration into the assimilation system and provides new ways for data assimilation of remote sensing and crop models.The main research contents and results of this dissertation are as follows:(1)Based on remote sensing data,regional statistical data,observations at agri-meteorological stations,and field measurements,MCMC calibration and posterior uncertainty analysis of the WOFOST(WOrld FOod Studies)crop growth model were carried out at different scales.Firstly,a likelihood function was constructed based on the phenological observations(dates of emergence,flowering,and maturity)of agri-meteorological stations,and the phenological parameters of the WOFOST model were calibrated using the MCMC method.Then,the likelihood function was constructed by county-level time-series LAI(Leaf area index)information from GLASS(Global land surface satellite)LAI products at 250-m spatial resolution and county-level statistical yield(for winter wheat estimation at county scale)or time-series GLASS LAI and field measured yield(for winter wheat estimation at field scale)and used the MCMC method to calibrate the key parameters of the WOFOST model and get their uncertainty estimation.The results show that the MCMC method can reasonably quantify the uncertainty of model parameters and their correlations based on different observational data and can provide quantitative estimates of model uncertainty for data assimilation systems.(2)The combination method of MCMC calibration posterior parameters and EnKF data assimilation algorithm was explored,constructing an EnKF assimilation estimation method that introduces a Bayesian posterior parameter set for yield estimation.Multi-resolution(10 km,250 m),large-scale(Henan and Hebei province)data assimilation-based yield estimation was carried out,and the results were compared with the classic EnKF assimilation method.The data assimilation results at 250-m spatial resolution were validated based on the field-measured yield.The results show that the overall accuracy reaches 92%(according to the MAPE(Mean Absolute Percentage Error)),the R~2reaches0.41,and the RMSE(Root Mean Square Error)reaches 686 kg/ha,which is significantly better than the classical EnKF results(R~2 is 0.22～0.38,RMSE is 708～928 kg/ha).Validation based on county statistical yield shows that the 10-km data assimilation results reach the overall accuracy of 88%～90%calculated by MAPE,with R~2 from 0.42 to 0.68,RMSE from 619 to 813 kg/ha,the overall results are better than the classical EnKF results with the average accuracy over the years is 82%to 88%calculated by MAPE.It shows that the proposed data assimilation method can quantify the uncertainty of the model more objectively and reasonably with better overall accuracy.(3)To improve the computational efficiency of data assimilation with high spatial resolution remote sensing data,the yield mapping method based on a Bayesian posterior prediction set and remote sensing LAI was explored,constructing a rapid yield estimation method suitable for high spatial resolution remote sensing data assimilating.Based on the county-level statistical yield in 2016,the field-measured crop growth information in 2017,and the remote sensing LAI retrieved by Sentinel-2 and Landsat 8,the regional yield at10-meter resolution was mapped for Hengshui City Hebei Province by this method.The results show that the yield estimation accuracy of the constructed method within the plot is R~2=0.29,RMSE=574kg/ha,and the yield estimation accuracy at the county scale is R~2=0.52,RMSE=442kg/ha.Meanwhile,the computational efficiency of the construction method is about 90 times that of the classical EnKF assimilation.