基于时间序列LAI和ET同化的冬小麦遥感估产方法比较

Comparison of remote sensing yield estimation methods for winter wheat based on assimilating time-sequence LAI and ET

为了评估同化时间序列叶面积指数(leaf area index,LAI)和蒸散发(evapotranspiration,ET)产品对冬小麦产量估测的有效性和适用性,该文选择陕西省关中平原冬小麦为研究对象,以SWAP为作物生长动态模型,利用冬小麦关键生育期的遥感观测和SWAP模拟LAI、ET趋势变化信息构建代价函数,以SCE-UA作为优化算法最小化代价函数,重新初始化SWAP模型中的出苗日期和灌溉量2个参数。重点比较了基于向量夹角和一阶差分2种代价函数的冬小麦单产估测精度。结果表明,同化MODIS LAI和ET后,冬小麦产量的估测精度比未同化精度(r=0.57,RMSE=1 192 kg/hm2)有显著提高,并且基于向量夹角代价函数法同化策略的单产估测精度(r=0.75,RMSE=494 kg/hm2)高于一阶差分代价函数法(r=0.73,RMSE=667 kg/hm2)的估测精度。该方法为其他区域的水分胁迫模式下遥感与作物模型双变量数据同化提供了参考。

Assimilating biophysical parameters derived from remote sensing into crop growth model is an important approach to improve performance of regional crop yield estimation. Currently, most researches adopt single remote sensing data source and single variable assimilation strategy, which cannot accurately reflect the interactive process among radiation, temperature and water, limiting the performance of data assimilation model. Leaf area index（LAI） and evapotranspiration（ET） are two key biophysical variables related to crop growth and grain yield. The study presents an assimilation framework to assimilate MODIS LAI product（MCD15A3） and MODIS ET product（MOD16A2） into the Soil-Water-Atmosphere-Plant（SWAP） model to improve the estimates of winter wheat yield at the regional scale. The spatial scale is one of the most challenging issues in the field of remote sensing, and the mismatching between remote sensing observations and state variables of crop model has an important impact on the performance of data assimilation model. MODIS LAI and ET products in 1 km scale are severely underestimated compared to the ground-based observations because of the mixed pixel effect and the heterogeneity within pixel, and hence the scale factors of 1 km MODIS products and the crop model＇s simulated parameters are totally different. So the direct assimilation of 1 km MODIS products would cause abnormal results. At present, there are two types of solutions to mitigate the scale issue; one is to scale down remote sensing parameters or scale up crop model＇s simulated variables, and the other is to assimilate the time series trend characteristics derived from remote sensing into crop model. In this study, two types of cost functions were constructed through comparing the generalized vector angle or first order difference of the observations and modeled LAI and ET time series trends during the growing season. Two key model parameters（i.e. irrigation water depth and emergence date） were selected as the reinitialized parameters needed to be optimized through minimizing the cost function using the SCE-UA optimization algorithm, and then the optimized parameters were input into the SWAP model for winter wheat yield estimation. Winter wheat yield assimilation estimation accuracy was evaluated for two cost functions（e.g., vector angle and first order difference） at field and regional scales. The results showed that yield estimation accuracy had been greatly improved with assimilation of LAI and ET trends than without assimilation. Furthermore, vector angle strategy（r=0.75, RMSE=494 kg/hm2） had achieved higher accuracy than first order difference（r=0.73, RMSE=667 kg/hm2）. In this study, equal weights were given to LAI and ET in the cost function. Giving different weights according to the errors of the LAI and ET data at different crop phenological stages would further improve the performance of data assimilation model. LAI and ET were selected as the assimilation variables in the cost function.Additional important state variables（e.g., soil moisture） that also closely related to grain yield should be incorporated into data assimilation framework to test the impacts to the crop yield. So, a more robust approach needs to be developed to simultaneously assimilate multiple biophysical variables（e.g., LAI, ET/PET, soil moisture）, and hybrid approaches, such as combining the use of En KF and 4DVar, would allow simultaneous estimates and updating of the model parameters and state variables, and would further improve crop yield estimation at field and regional scales.