A Study On The Uncertainty Of Regional Winter Wheat Growth Simulation From A Crop Model Using Remote Sensing Data Assimilation

For sustainable agriculture management and to assess national food security, it is important to determine regional crop yield information in a timely and accurate manner. Data assimilation approach that integrate crop growth models with remote sensing information have been proposed as a hot spot and developed as a potential approach to improve the accuracy of regional crop yield. But this simulation process is complex. Some uncertain factors that may result in errors in the crop model data assimilation framework should be considered. Analysis and study on these uncertainty issues will certainly improve the validity of crop model data assimilation, and then promote the developments of agricultural remote sensing monitoring capacity.In this study, we mainly focus on the uncertainties of crop growth simulation and yield estimates due to different assimilation strategies and multi-source errors, such as initial conditions, simulating process, data assimilation algorithms and internal parameters, remotely sensed observations, weather data, the model itself, and spatio-temporal scales. First, based on the localizing of CERES-Wheat model, the uncertainties due to errors that were derived from the initial conditions and/or ‘imperfect’ simulations were the research emphasis. Second, we compared the simulated results with particle filter(PF) and the proper orthogonal decomposition–based ensemble four dimensional variational(POD4DVAR) assimilation schemes, and analyzed the uncertainties of the particle/ensemble dimensions and perturbed variance. Then, remote sensed LAI were inverted with the PROSAIL model. At last, a series of regional winter wheat LAI in Hengshui were assimilated into the CERES-Wheat model based on the POD4DVAR-based strategy, and the uncertainties of the meteorological forcing data, observational errors, and temporal-spatial scales were also analyzed. The main contents and results are as follows:(1) Three scenarios with optimization of initial conditions, PF assimilation, and synchronous optimization and PF assimilation were used for crop growth simulation and yield estimates. The results showed that these three scenarios all can improve the accuracies of simulated LAI. The first scenario with optimization of initial conditions could not improve the accuracies of yield estimates, while other two scenarios with considering the uncertainties of simulations and external observations, improved the accuracies. Especially, the third scenarios also considered the initial conditions, and was the relatively optimal scheme with a relative error(RE) of 6.00%, and a root mean square error(RMSE) of 544kg/ha. Then the effects of observation errors and assimilation phenological stages were also analyzed, and the results were showed that with the increasing of observational errors, the accuracies of assimilation simulations were decreased. It is also conducted that assimilation simulations can be improve significantly when assimilating of observation with the single growth period of booting, heading and jointing stages, and then also the same with assimilating of observations with early and middle phenological phases.(2) Comparing of simulated results with PF and POD4 DVAR assimilation algorithms, it is concluded that the two schemes both can improve the accuracies of assimilating simulations, of which POD4 DVAR received relative better results with the RE of 5.65% and RMSE of 523 kg/ha. The study also found that with the increasing of particle/ensemble dimensions, the improvements were not significantly. But the computational cost increased above 8 times. As perturbed variance of particles increased, it is not significant improved the simulations. Therefore, the assimilation algorithms, disturbed particle/ensemble dimensions and variance of particle/ensemble should be determined to trade-off the simulated accuracy and computed cost for regional assimilation simulations of winter wheat.(3) Using simultaneously collected remote sensing data and field measurements, this study firstly assessed the consistency and applicability of China high-resolution earth observation system satellite 1(GF-1) wide field of view camera(WFV), environment and disaster monitoring and forecasting satellite(HJ-1) Charge Coupled Device(CCD), and Landsat-8 Operational Land Imager(OLI) data for estimating the leaf area index(LAI) of winter wheat via reflectance and vegetation indices. The LAI results from three sensors have better consistency and applicability, with PROSAIL model. Then the time-series LAIs of winter wheat were inversed, and the validations with different periods and phenological stages were also performed. The overall RE and RMSE were 5.72% and 0.26 at different periods, and the RE and RMSE were 9.44% and 0.39 at different phenological stages. The results met the needs of research on regional wheat growth simulation data assimilation.(4) A series of regional winter wheat LAI in Hengshui were assimilated into the CERES-Wheat model based on the POD4DVAR-based strategy, and the acceptable estimates of the winter wheat yield were obtained with the RE and RMSE were 8.32% and 452kg/ha. The effects of the meteorological forcing data were also analyzed, and the yield estimation by the driven data from single meteorological station, cannot reflect the changes of actual yield by the actual weather conditions. The results also showed that with the increasing of remote sensed LAI errors, the accuracies of yield estimates were decreased. But the deteriorations were not significantly, and the POD4DVAR-based strategy can eased part affection of remote sensed LAI errors. In is also conducted that with the increasing of assimilation observational frequency, the performance of yield estimates improved. In contrast, an acceptable estimate of crop yield could also be achieved by assimilating fewer observations during the booting, heading and jointing stages, so as to the early and middle phenological stages. The accuracy of the regional crop yield estimates generally decreased as the spatial resolution of the LAI maps decreased. Conversely, the computation times were lower for the assimilation of observations with lower frequencies and spatial scales. Thus, it is important to consider reasonable temporal-spatial scales and assimilation stages to obtain tradeoffs between accuracy and computation time, especially in operational systems used for regional crop yield estimates.