SCE标定结合EnKF同化遥感和WOFOST模型模拟冬小麦时序LAI

WOFOST(world food studies)模型可用于模拟冬小麦全生育期内的时序叶面积指数(leaf area index, LAI),各器官生物量以及最终产量,对冬小麦的长势监测与产量预估有着重要意义。但将WOFOST模型用于中国具体区域的冬小麦生长模拟时,存在着参数定标困难、模拟结果不够准确等严重问题。目前对该模型的定标大多依靠研究者的经验进行,虽已总结出了一套从标定到模拟应用的研究方法,但在区域模拟时仍然存在很多问题。为此,该文以较易获取的LAI为参考指标,结合潜在生长水平模式下的WOFOST模型在衡水地区的应用,提出了一种"区域优化标定,像元同化修正"的研究方法:首先在区域尺度上对WOFOST模型进行优化标定,利用扩展傅里叶幅度灵敏度检验法(extend fourier amplitude sensitivity test, EFAST)分析模型各个参数的敏感性,在此基础上选择了可以迅速找到全局最优解的SCE(shuffled complex evolution)算法对总敏感度最高的5个参数进行优化,并将优化前后的时序LAI曲线进行对比;其次运用第一步确定的模型最优参数,在对区域内每个像元进行模拟时,结合Sentinel-2卫星数据反演所得的各个像元LAI,利用集合卡尔曼滤波(ensemble kalman filter, EnKF)在像元尺度上对LAI进行同化修正,并结合采样点的2次实测LAI数据对同化所得结果进行验证。试验发现,优化标定后的WOFOST模型模拟所得LAI曲线更接近所给的LAI真值,在此基础上结合数据同化模拟得出的衡水地区每个像元LAI的R2达到0.87,RMSE仅为0.62。因此,与原来只能通过经验进行定标的方法相比,该方法有效地解决了WOFOST模型在具体应用中亟待解决的复杂标定问题,并且结合同化修正有效地提高了模型在各个像元的模拟精度,R2由0.70～0.83提升至了0.87,RMSE由0.89～1.36降低至了0.62。同时该文也提供了从模型标定到具体模拟整个过程中各个环节的思路与方法,有利于促进WOFOST模型在区域尺度上的应用。

Time-series LAI simulation of winter wheat based on WOFOST model calibrated by SCE and assimilated by EnKF

WOFOST (world food studies) model can be used to simulate time-series LAI (leaf area index), the organs'' biomass, and the yields of winter wheat. Therefore, it is meaningful for the growth monitoring and production prediction of winter wheat. So far, the calibration of WOFOST usually relies on researches'' experience, which brings many problems while using the model in a specific area. As a result, we focus on the calibration problem and try to improve the accuracy of the simulated results in this paper. The potential production WOFOST was analyzed and LAI was chosen as the measure index because it was easy to obtain. In this study, we selected Hengshui as the study area, and two field experiments were carried out in this area during two different periods. One period was from 2017-03-29 to 2017-04-01 and the other was from 2017-05-04 to 2017-05-06. It was divided into 11 sampling areas and 5 sampling points in every area were obtained to measure the LAI, so we got approximately 110 measured data totally. A method called ''Calibrating in area by optimization and correcting at pixel by assimilation'' was presented in this paper. Firstly, calibrating WOFOST model in local area: The weather data including sunshine duration data and the maximum and minimum air temperature data every day were used to run the WOFOST model. The data were from Nangong National Weather Station and can be downloaded in National Meteorological Information Center. Then the sensitivity of model parameters can be analyzed with EFAST (extend fourier amplitude sensitivity test) and the 5 most sensitive parameters were selected to optimize the model. It was worthwhile to note that there were different indices to evaluate the sensitivity of every parameter, such as main effect, interaction, and total effect, and the total effect was considered as the most important index in this study. As for the optimization, the SCE (shuffled complex evolution) algorithm was used which could find the global optimal solution fastly. It can solve the initial value dependence problem and local convergence problem which might exist in other optimization algorithms such as MCMC (Markov Chain Monte Carlo). In order to proof that the optimization was valid, the time-series LAI curves simulated were compared by WOFOST before and after optimization with SCE with measured values. It turned out the model after optimization was much more appropriate to simulate the growth of winter wheat in study area. Secondly, assimilating the model in every pixel in the study area: We interpolated weather data from 21 National Weather Stations in Hebei Province in order to run WOFOST in every pixel. Based on this, EnKF (Ensemble Kalman Filter) was used to assimilate LAI in every pixel with the remote sensing data from Sentinel-2. As a result, we could get the time-series LAI curve at every pixel. The LAI curve at point HS01 was illustrated and it was obvious that assimilation made a difference in the simulation. Additionally, the simulated LAI distribution maps were illustrated in Hengshui at date of 2017-03-30 and 2017-05-05. And the simulated LAI values of the pixels according to the sampling points were extracted. By comparing the simulated LAI with measured LAI, we found that R2 was increased from 0.70-0.83 to 0.87 and RMSE was decreased from 0.89-1.36 to 0.62. Therefore, the method proposed in this study solved the calibration problem and improved the accuracy of time-series LAI simulated compared with other studies. In addition, we provided specific theories and methods in every stage from calibration to application. It contributed to the application of WOFOST in our country.