Winter Wheat Yield Estimation Based On Multi-variables And Data Assimilation Algorithms

Accurate crop growth monitoring and yield estimation and prediction are the guarantee for national food security.The development of the remote sensing technology offers a new direction for timely,dynamic and up-to-date crop yield estimation and prediction over large areas.Such as,the remotely sensed data can be utilized in conjunction with crop growth models to estimate regional grain yields on the basis of the data assimilation approach,which takes into consideration the mechanism of the crop growth as well as the effect of environmental factors on crop growth and development.Also this approach can solve the difficulty in regional parameters acquisition of crop growth models.In this study the Guanzhong Plain in Shaanxi provience was selected as the study area,and the Landsat-5,Landsat-7 and Landsat-8 images at the main wheat growth stages were obtained for retrieving normalized difference vegetation index(NDVI)and vegetation temperature condition index(VTCI).Then the empirical regression model between NDVI and field-measured leaf area index(LAI)and that between NDVI and measured aboveground dry biomass(?)at the sites were used for estimating regional LAI and ? values respectively.Also the regression model between VTCI and soil moisture at the depth of 0-20 cm(?)was applied for estimating regional ? values.In addition,the genetic parameters of the CERES-Wheat model were calibrated by comparing the simulated state variables,grain yields and harvest date with the field measurements collected during the winter wheat growing seasons.Then the simulation performances of the calibrated CERES-Wheat model were evaluated using the field measurements as well.Results indicated that(1)the mean relative error(MRE)between the simulated and measured ? and that between the simulated and measured ? were both lower than 10%;(2)the MRE between the simulated and measured yields was lower than 15%;and(3)the difference between the simulated and surveyed harvest date was less than 4 days.Thus the simulation accuracies of the CERES-Wheat model were high.The LAI,? and ? simulated by the CERES-Wheat model were assimilated with the state variables derived from Landsat data,for obtaining the daily assimilated LAI,? and ? values using the four-dimensional variational(4DVAR),ensemble Kalman filter(EnKF)and particle filter(PF)algorithms respectively.Then the assimilated variables were compared with both the simulated variables and the field-measurements for validating the accuracies of the assimilation performances.Results showed that the assimilated LAI,? and ? trajectories followed the phenological characteristics of the simulated variables,and were influenced by the remotely sensed variables as well.The accuracies of the PF algorithm in assimilating LAI and ? were higher than the accuracies of the EnKF and 4DVAR algorithms in assimilating LAI and ?.Meanwhile the assimilated LA1 and ? trajectories by the PF algorithm effectively reflected the phenological characteristics of the simulated variables.The EnKF algorithm performed well in terms of reflecting the temporal variation characteristics of the simulated LAI and ?,however,the RMSEs between the assimilated LAI and ? values and the field-measurements were higher.In addition,the accuracies of the assimilated LAI and ? based on the 4DVAR algorithm were high,but the application of assimilation time window would result in the lack of assimilated data.The linear regression analyses,with the assimilated state variables as the independent variables and the measured yields as the dependent variable,were carried out for establishing the yield estimation models for each wheat growth stage.Then the weights of the yield estimation models for each growth stage,which were obtained using a combination entropy forecasting method,were used to establish the combination yield estimation models,and the accuracies of the yield estimation models were validated using the field-measured yields.Results showed the accuracies of the yield estimation models based on the PF algorithm were higher than the accuracies of the yield estimation models based on the EnKF and 4DVAR algorithms.The variables,which were highly correlated with the measured yields,were selected as the optimal-assimilation variables for the corresponding stage.The yield estimation models were established based on three different assimilation strategies:(1)the assimilation of LAI and ?(or LAI and ?,or ? and ?);(2)the assimilation of LAI,? and ? simultaneously;and(3)the assimilation of the optimal-assimilation variables at each wheat growth stage.Results showed that? and ? were chosen as the optimal-assimilation variables for the green-up and milk stages respectively,and LAI,? and ? were chosen as the optimal-assimilation variables simultaneously for the jointing and heading-filling stages.The accuracy of the yield estimation model established by assimilating the optimal-assimilation variables(R2=0.81,RMSE=317.85 kg·ha-1)was higher than the accuracy of the model established by assimilating LAI,,? and ? simultaneously(R2=0.76,RMSE=348.64 kg·ha-1).In addition,the accuracy of the yield estimation model established by assimilating LAI,? and ? was higher than the accuracy of the model established by assimilating bivariables.Therefore the assimilation of highly yield-related state variables at each wheat growth stage provides a reliable and promising method for improving crop yield estimates.The Guanzhong Plain includes both irrigated and rain-fed farmlands,and the yield estimation model established separately for the irrigated and rain-fed fields resulted in better accuracy(R2=0.85,RMSE=287.55 kg·ha-1)than the model established for both the irrigated and rain-fed fields.Then the yield estimation model for the irrigated and rain-fed farmlands respectively were used to estimate the regional winter wheat yields in 2007-2008 and 2013-2016,and the spatial distribution characteristics of wheat yields were analyzed as well.Results showed wheat planting densities in the middle and weatern regions of the Guanzhong Plain were high while those in the eastern and northern regions were relatively low.The average wheat yields of the eastern and northern regions were also lower than the wheat yields of the middle and weatern regions.