粒子滤波同化方法在CERES-Wheat作物模型估产中的应用

为验证粒子滤波同化算法在作物模型估产应用中的可行性和精度,应用该算法构建了CERES-Wheat(crop environment resource synthesis for wheat)作物模型同化系统,并应用地面观测数据研究了同化系统的估产能力以及粒子扰动维数和方差对同化结果和效率的影响。研究结果表明,构建的作物模型同化系统能够利用作物关键生育期内观测LAI数据,较好地校正模型状态轨迹,显著提高作物产量模拟预测精度。同化前后冬小麦产量模拟结果与实测产量间的决定系数由0.68增加为0.83,归一化均方根误差由4.93%减小为3.4%,相对误差由4.15%减小为2.93%。粒子扰动维数和方差同化试验结果显示,粒子维数由50增加为250时,同化估产精度无显著改善,但计算代价增加5倍;随粒子扰动方差增加,估产归一化均方根误差和相对误差均呈增加趋势,两者的平均增加幅度分别为0.32%和0.26%。因此,作物模型同化系统业务化应用时需折中考虑估产精度和计算代价设置合适的粒子扰动维数和方差。该文为进一步利用多源卫星遥感数据监测区域作物长势和估算产量等同化研究和应用提供参考。

Estimation of crop yield using CERES-Wheat model based on particle filter data assimilation method

In order to validate the feasibility of particle filter algorithm for crop yield estimation using crop growth model as well as accuracy, the data assimilation system of CERES-Wheat (crop environment resource synthesis for wheat) model was integrated. The ground observation was employed to perform an experiment to test the capability of data assimilation system for crop yield estimation. The effect of particle dimension and perturbed variance on accuracy and efficiency was also discussed. The results showed that CERES-Wheat data assimilation system based on particle filter could correctly adjust the trajectory of model states and significantly improve the precision of forecasted crop yield. The coefficient of determination R2 between simulated and observed winter wheat yields before and after performing PF assimilation scheme increased from 0.68 to 0.83, normalized root mean square error (NRMSE) from 4.93% to 3.40%, and relative error (RE) from 4.15% to 2.93%. For the trials of perturbed particle size and variance, there were no significant improvement when particle size increased from 50 to 250, but the computation time increased by 5 times. As perturbed variance of particles increased, NRMSE and RE increased by 0.32% and 0.26%, respectively. Therefore, the disturbed dimensions and variance of particles should be determined based on a compromised consideration of both the simulated precision and computed cost. This study provides a valuable reference for monitoring crop growth and yield estimation on regional scale using multi-sources remote sensing data.