Estimation of wheat yield based on crop and remote sensing assimilation models

Abstract: Data assimilation (DA) has been recognized as a promising approach for regional crop growth monitoring and yield estimation. The widely used DA method, ensemble Kalman filter (EnKF), holds the assumption that the involved probability density functions (PDFs) are Gaussian, and the evolution of the filter can be governed only by its second-order characteristics, leading to a significant loss of information. In comparison with the EnKF, the particle filter (PF) has no restrictive assumption regarding the forms of the PDFs, and thus can be applied to any nonlinear and non-Gaussian systems. Different researchers have used leaf area index (LAI), vegetation indices and soil moisture as the state variables in agricultural data-assimilation systems for estimating crop yields. However, the assimilation of variables that are not very important for crop yields (e.g., LAI at the maturity stage) may decrease the accuracy of yield estimations. Conversely, assimilating highly yield-related variables is important for improving yield estimates. To improve winter wheat yield esimation in the Guanzhong Plain, China and determine whether assimilating highly yield-related variables at each wheat growth stage improved the accuracy of the yield estimation, daily LAI, soil moisture (0-20 cm) and aboveground dry biomass simulated by the CERES-Wheat model were assimilated from the LAI, soil moisture and biomass retrieved from Landsat data using the PF algorithm, for obtaining daily assimilated LAI, soil moisture and biomass values. Then, the daily assimilated LAI and biomass values during the growth stages of winter wheat, including the green-up, jointing, heading-filling and milk stages, were accumulated to obtain the accumulated LAI and biomass values. Linear regression analyses were performed to examine the relationships between accumulated LAI, accumulated biomass or assimilated soil moisture and the field-measured yields respectively for determining the optimal-assimilation variables. The results showed that the PF algorithm combined the remotely sensed LAI, soil moisture and biomass values with the phenological characteristics of simulated LAI, soil moisture and biomass trajectories, which improved the daily LAI, soil moisture and biomass estimation. The field measurements for the sampling sites were compared with the assimilated and simulated LAI, biomass and soil moisture. The RMSE of 0.61 m2/m2 and 790.65 kg/hm2 of the assimilated LAI and biomassvalues were 0.25 m2/m2 and 154.21 kg/hm2 lower than those of the simulated LAI and biomass values. Similarly, the RMSE of 0.017 mm3/mm3 of the assimilated soil moisture value was 0.012 mm3/mm3 lower than that of the simulated soil moisture value. At the green-up stage, the linear correlation between the assimilated soil moisture and the field-measured yields was higher than those between the accumulated LAI and the yields or between the accumulated biomass and the yields, that soil moisture was chosen as the optimal-assimilation variable for the green-up stage. At the jointing and heading-filling stages, the accumulated LAI, accumulated biomass or assimilated soil moisture were all highly correlated to the yields, respectively, and thus all of them were chosen as the optimal-assimilation variables for the 2 stage. In addition, biomass was selected as the optimal-assimilation variable for the milk stage. The optimal-assimilation yield estimation model, established based on the optimal-assimilation variables at each growth stage, achieved better estimation accuracy for wheat yields (R2=0.91, RMSE=207.76 kg/hm2) than the yield estimation model established based on the assimilation of LAI, soil moisture and biomass simultaneously (R2=0.84, RMSE=281.69 kg/hm2). Moreover, the yield estimation accuracy by assimilating LAI, soil moisture and biomass was higher than that by assimilating LAI and soil moisture (or soil moisture and biomass, or LAI and soil moisture). Therefore, assimilating highly yield-related variables at each crop growth stage provides reliable and promising methods for improving crop yield estimates.