Winter wheat yield estimation based on assimilation method combined with 4DVAR and EnKF

Abstract: The CERES-Wheat model is used to simulate leaf area index (LAI) of winter wheat for reflecting accurately the growth of winter wheat and estimating the yield. But it is difficult to simulate wheat yield at a large area with CERES-Wheat model because of the lack of regional input parameters. Data assimilation algorithm is an efficient method to combine crop growth model and remote sensing data and it solves the shortcoming of CERES-Wheat model by taking advantage of macro of remote sensing data. NASA's Moderate Resolution Imaging Spectroradiometer (MODIS) data is normally used in assimilation with crop growth model for yield estimation; but the spatial resolution of MODIS is low, and the estimation accuracy of crop will be reduced by the mixed pixel problem, as the farmland is small. Landsat remotely sensed data with a spatial resolution of 30 m would be helpful to make crop yield estimation at field scale in China, and TM and ETM+ remotely sensed data were both used in assimilation for solving the disadvantage of low temporal resolution of Landsat. The CERES-Wheat model was used to simulate LAI of the whole growth period of winter wheat in Guanzhong Plain of Shaanxi province. The assimilation of simulated LAI with LAI retrieved from TM and ETM+ data was carried out in eight typical sampling sites by using two data assimilation approaches, the four-dimensional variational (4DVAR) and ensemble Kalman filter (EnKF). The assimilated LAI image of the whole study area was achieved by employing the linear correlation model between remotely sensed LAI and assimilated LAI of the eight sampling sites. After establishing the assimilation system, the remotely sensed LAI and the simulated ones of the eight sampling sites were used to test the two assimilation approaches. The result showed that the assimilated LAI values of the sites were more accurate and closer to the real ones after combining the advantages of both remotely sensed LAI and simulated LAI. In order to utilize a more accurate assimilation algorithm for estimating yield of winter wheat in Guanzhong Plain, the 4DVAR and EnKF approaches were compared and analyzed for both assimilated LAI values of the sites and assimilated LAI images of the two assimilation approaches. The root mean square error (RMSE) between EnKF-LAI values and measured LAI values of the sites was 0.41, while the RMSE of 4DVAR-LAI values was 0.55, so EnKF-LAI values of the eight sites were closer to the measured ones than those of the 4DVAR-LAI. After comparing assimilated LAI images from 4DVAR and EnKF approaches, it concluded that the EnKF-LAI images were more in line with spatial distribution characteristics of winter wheat's LAI in Guanzhong Plain, and the EnKF algorithm was an appropriated approach for assimilating LAI. The EnKF-LAI images of 4 main growth stages of winter wheat including the reviving stage, jointing stage, heading-filling stage and dough stage were used to estimate winter wheat yield of the whole plain by constructing winter wheat yield estimation model. By comparing with the measured yields of winter wheat in the crop year of 2007-2008, the relative errors (RE) of the estimated yields of the eight sites were from 2.05% to 14.57% with an average of 9.15%, and the RMSE between the estimated and measured yields was 596.7 kg/hm2. While the RE between the simulated and measured yields was from 8.80% to 36.61% with an average of 22.13%, and the RMSE was 1699.0 kg/hm2. In addition, wheat yields in the crop year of 2013-2014 were used for further validation of the winter wheat yield estimation model. Compared with the RE between the simulated and measured yields, the RE between the estimated and measured yields was decreased by 0.57%-9.30%, with an average relative error reduction of 3.89% in the 15 sampling sites; and the estimated accuracies in the 8 sampling sites are larger than 94%. These show that the estimated yields of the sampling sites are close to the measured ones, and the accuracy of the yield estimation model is obviously improved.