顺序同化不同时空分辨率LAI的冬小麦估产对比研究

Comparison of Winter Wheat Yield Estimation by Sequential Assimilation of Different Spatio-temporal Resolution Remotely Sensed LAI Datasets

选择PyWOFOST模型为动态模型,以叶面积指数(LAI)为状态变量,遥感LAI为观测值,采用集合卡尔曼滤波(En KF)同化算法,研发了一种遥感LAI与作物模型同化的区域冬小麦产量估测系统。为消除云的污染,采用Savitzky-Golay(S-G)滤波算法重构时间序列MODIS LAI;通过构建地面观测LAI与3个关键物候期Landsat TM植被指数回归统计模型,获得区域TM LAI;通过融合3个关键物候期的TM LAI与时间序列S-G MODIS LAI,生成尺度转换LAI。对比分析3种不同时空分辨率的遥感LAI的同化精度,研究结果表明,同化尺度转换LAI获得了最高的同化精度,与官方县域统计产量相比,在潜在模式下,决定系数由同化前的0.24提高到0.47,均方根误差由602kg/hm2下降到478 kg/hm2。结果表明,遥感观测与作物模型的尺度调整对提高冬小麦同化模型精度具有重要作用,遥感LAI与作物模型的En KF同化方法是一种有效的区域作物产量估测方法。

Data assimilation method combines with remotely sensed data and crop growth model has become an important hotspot in crop yield forecasting. PyWOFOST model and remotely sensed LAI were respectively selected as the crop growth model and observations to construct a regional winter wheat yield forecasting scheme with EnKF algorithm. To eliminate cloud contamination, a Savitzky - Golay （ S - G） filtering algorithm was applied to the MODIS LAI products to obtain filtered LAIs. Regression models between field-measured LAI and Landsat TM vegetation indices were established and multi-temporal TM LAIs was derived. The TM LAI with time series of MODIS LAI was integrated to generate scale-adjusted LAI. Compared the assimilation accuracy using these three different spatio-temporal resolution remotely sensed data, validation results demonstrated that assimilating the scale-adjusted LAI achieved the best prediction accuracy, in potential mode, the determination coefficient （R2） increased from 0.24 which without assimilation to 0.47 and RMSE decreased from 602 kg/hm2 to 478 kg/hm2 at county level compared to the official statistical yield data. Our results indicated that the scale adjustment between remotely sensed observation and crop model greatly improved the accuracy of winter wheat yield forecasting. The assimilation of remotely sensed data into crop growth model with EnKF can provide a reliable approach for regional crop yield estimation.