Crop model data assimilation with particle filter for yield prediction using leaf area index of different temporal scales

Crop yield prediction is a significant component of national food security assessment and food policy making. The data assimilation method which combined crop growth model and multi-source observed data has been proven to be the most effective method to simulate the crop growth process and predict crop yields. Based on the observed LAI, the time-series LAI data sets were obtained by using the smooth unequally spaced interpolation. With the support of amount of related data, the crop model of Crop Environment Resource Synthesis (CERES)-Wheat can be used to simulate the growth of the winter wheat, and the particle filtering-based data assimilation strategy also could be implemented to improve the CERES-Wheat simulation process. Through the assimilation strategy, the incorporation between the observed and modeled LAIs was performed under the dynamic framework of the winter wheat growth process such that the LAI simulation was sequentially optimized, which led to optimal yield estimation. Comparing the measured the data with the assimilation and no assimilation results, it obvious showed that the yield estimations were dramatically improved after assimilation. Then considering the influence of the observed leaf area index on the assimilation process at multi-temporal scale, this paper try to analysis the uncertainty of the assimilation results from three aspects of temporal scale: First, Assimilating the observed LAI data at five kinds of temporal resolution (2-day, 4-day, 8-day, 16-day, 32-day), the results showed that the accuracy of the yield estimation tended to significantly improve as the temporal resolution of assimilating LAI increased from 32-day to 2-day. In addition, as the increasing of the temporal resolution, the computing time increased dramatically. Second, the observed LAI data was applied to assimilation with the four different stages of crop growth (green-jointing, jointing-heading, heading-mature, green-mature). The results showed that the accuracy of the yield estimation by assimilating the LAI at the stages of the green-mature (the whole growth phase) is the highest, while the accuracy during the stages from the green-jointing is the lowest. And also the accuracy during the stages from jointing-heading is the second highest. Third, the observed LAI data was used to assimilation with synthesizing the temporal resolution and the stages of crop growth (three kinds of temporal resolution (2, 4, 8 day) and four stages of crop growth). The results showed that on the premise of the optimal efficiency and accuracy, the optimal plan is to select observed LAI by using the temporal resolution of 8 day during the stages from jointing-heading. This study will be a better influence on decreasing the calculation time and improving the effect for the crop model data assimilation.