Winter Wheat Yield Estimation Based On Assimilated Sentinel Multi-source Data With Crop Growth Model

Crop growth monitoring and accurate yield forecasting and estimation are crucial for agricultural management and national food security,especially in the wake of COVID-19 causing long-term impact and the volatile international situation.Data assimilation can substantially help improve the simulation accuracy of the crop model at a regional scale by integrating the remote sensing information on the spatial-temporal distribution of crop canopy,soil and atmospheric parameters at a regional scale into the process-based crop model,and adjusting the state variables of the model simulation or optimizing the initial state parameters.In this thesis,winter wheat growing areas in three counties in southern Shanxi Province were selected as the study area to improve the accuracy and efficiency of yield estimation based on sensing information and crop modeling from four aspects including accuracy and frequency of remote sensing information,assimilation algorithm,water stress partitioning and assimilation estimation strategy.The enhanced temporal adaptive reflectance fusion model(ESTARFM)algorithm was used to fuse the data of 6 Sentienl-2 and 15 Sentienl-3 optical images with cloud cover <5% during the critical growth stage(March 1-May 30)of winter wheat in the study area,and a total of 15 fused data with 20 m spatial resolution were obtained.The accuracy of ESTARFM data on March 12 and April 16 was examined against the corresponding bands of Sentienl-2 on March 12 and April 17,and the correlation coefficient R2 values ranged from 0.6 to 0.85,and the RMSE values ranged from 0.09 to 0.13.The spatial resolution and distribution of spatial information of ESTARFM data are basically consistent with Sentienl-2,which can clearly express the boundaries of rivers and construction land and plots of cultivated land,and can obviously reflect the crop change over time.Accuracy(RMSE)of LAI inversed from ESTARFM fusion data was 0.9742 m2 /m2,with a temporal frequency of 6 d in the critical growth stage of winter wheat,and the Accuracy(RMSE)of LAI inversed from Sentienl-2 images was 0.9955 m2/m2,but with a temporal frequency of 15 d in the critical growth stage of winter wheat.Therefore,ESTARFM fusion data can provide LAI with high frequency and high accuracy for crop model assimilation.4DVAR,En KF and PF algorithms were applied to assimilate LAI inversed from ESTARFM fusion data,soil moisture inversed from Sentienl-1 microwave data,and the LAI and soil moisture obtained from CERESWheat model simulation,thus the LAI and soil moisture assimilated by 4DVAR,En KF and PF were obtained as follows: 4DVAR_ESTARFM_LAI,En KF_ESTARFM_LAI,PF_ESTARFM_LAI and 4DVAR_θ,En KF_θ,PF_θ,respectively.By validating against the measured LAI and soil moisture,it was found that the R2 of the three algorithms assimilated LAIs and model simulated LAI were 0.7522,0.7357,0.7523 and0.6699,respectively,and the RMSEs were 0.6402,0.6549,0.6642 and 1.2237 m2/m2,respectively.We found that the R2 of assimilated LAI by the three algorithms and RMSE were rather similar,all higher than that of model simulated LAI,and RMSE was on average 0.6 m2/m2 lower than that of model simulated LAI.The R2 of soil moisture assimilated by the three algorithms and model simulated soil moisture were 0.7036,0.6609,0.6635,and 0.6461,and RMSE were 0.0201,0.0296,0.02358 and 0.0562 cm3/cm3,respectively.We found that the R2 of soil moisture assimilated by all three algorithms was higher than that of model-simulated soil moisture,and the RMSE was lower than that of model-simulated soil moisture by 0.03 cm3/cm3 on average.This indicates that the LAI and soil moisture assimilated by all three algorithms can better combine the mechanistic characteristics of CERES-Wheat model simulation and the advantages of spatial continuity and temporal dynamics of Sentinel multi-source data,thus improving the accuracy of LAI and soil moisture.Among them,the accuracy of 4DVAR_ESTARFM_LAI,4DVAR_θ is slightly higher than that of En KF_ESTARFM_LAI,En KF_θ and PF_ESTARFM_LAI,PF_θ,respectively.The application of 4DVAR,En KF and PF algorithms to assimilate the LAI inversed from Sentinel-2images with the LAI inversed from CERES-Wheat model simulation enabled the obtaining of4DVAR_S2_LAI,En KF_S2_LAI and PF_S2_LAI.The R2 of 4DVAR_ESTARFM_LAI,En KF_ESTARFM_LAI and PF_ESTARFM_LAI were 0.016,0.008,and 0.0284 higher than that of4DVAR_S2_LAI,En KF_S2_LAI,and PF_S2_LAI,respectively,with RMSE of 0.1857,0.1292,and 0.1751m2/m2,respectively.This indicates that the ESTARFM fusion data performs better in inversion of LAI than Sentinel-2.The winter wheat growing areas were divided into water-stressed areas(arid areas)and non-waterstressed areas(irrigated areas)according to the different water stress conditions the winter wheat subjected to,and different assimilation estimation strategies were used for different water stress conditions:assimilation of LAI(single variable)or assimilation of LAI and soil moisture(double variant)to find a highly accurate and efficient assimilation estimation solution suitable for different water stress conditions.The results showed that for non-water-stressed areas,single variable LAI is enough for highly accurate yield estimation(RMSE=407.01 Kg/ha),However,in water-stressed areas,both LAI and soil moisture(double variant)need to be assimilated to correct both canopy growth and soil water balance processes to achieve the highest accuracy(RMSE=424.75 Kg/ha).It can achieve higher estimation accuracy by dividing winter wheat growing areas into irrigated and arid areas than yield estimation of the areas as a whole.The spatial distribution of winter wheat yields in the study area estimated by assimilation based on such division was consistent with the spatial distribution of the topographic location,irrigation conditions,and the cultivated land quality grade.