基于PEST的土壤-作物系统模型参数优化及灵敏度分析

Parameter optimization and sensitivity analysis of soil-crop system model using PEST

农业生产管理系统模型输入参数多,参数率定过程十分耗时费力,大大限制了其推广应用。该研究以华北平原2 a的冬小麦-夏玉米田间试验观测数据为基础,使用PEST(parameter estimation)参数自动优化工具对土壤-作物-大气系统水热碳氮过程藕合模型(soil water heat carbon and nitrogen simulator,WHCNS)的土壤水力学参数、氮素转化参数和作物遗传参数进行自动寻优,同时计算分析模型参数的相对综合敏感度,并将优化结果与土壤实测水力学参数和试错法的模拟结果进行比较。参数敏感度分析结果表明,18个模型参数的相对综合敏感度较高,其中土壤水力学参数普遍具有较高的敏感度,以饱和含水率敏感度最高;作物参数中,作物生长发育总积温和最大比叶面积具有较高的综合敏感度;而氮素转化参数的敏感度远低于土壤水力学参数和作物参数。评价模型模拟效果的统计性指标(均方根误差、模型效率系数和一致性指数)表明,PEST法比实测水力学参数的模拟精度有所提高,其中土壤含水率、土壤硝态氮含量、作物产量和叶面积指数的均方根误差分别降低了61.8%、23.5%、73.6%和23.3%。同时PEST法比试错法对土壤水分和作物产量的模拟精度也有较大提高,但对土壤氮素和叶面积指数的模拟精度提高不明显。由于该方法大大节约了模型校准时间,在较短的时间内获得了明显高于试错法的模拟精度,因此PEST软件在WHCNS模型参数自动优化中是一个值得推广的工具。

Agricultural production management system models usually require numerous input parameters and the calibration and validation of the parameters are time-consuming, which significantly limit the use of models. This study aimed at improving the efficiency and accuracy of a soil-crop system model （soil water heat carbon and nitrogen simulator, WHCNS） using a model-independent optimization tools （parameter estimation, PEST） and data from field experiments. A two-year field experiment was conducted from October 2009 to October 2011 in Tai＇an City, Shandong Province in North China Plain. The crop rotation was winter wheat-summer maize, and three fields with high, middle and low productivity levels based on the wheat yields （named T1, T2 and T3 treatments, respectively） were selected to test the WHCNS model. The dynamics of soil water content and soil nitrate concentration in different soil depths were monitored, crop dry matter and leaf area index at the key crop growth stages and yield data were measured. PEST was used to optimize model parameters and to calculate the relative composite sensitivity （RCS） of each input parameter for WHCNS model. The optimization parameters involved the majority modules of the model, such as soil water dynamic, nitrogen transformation and crop growth. The objective function of the optimization model were consist of four different groups of field data, including soil water content, soil nitrate concentration, crop yield and leaf area index. And the inverse solution was obtained through minimizing the object function using PEST program base on Gauss-Marquardt-Levevberg algorithm. The results of PEST were then compared with the simulations based on measured soil hydraulic parameters and the trial-and-error method. The statistical analysis （root mean square error, model efficiency, and agreement index） indicated that the PEST optimization method provided better accuracy and efficiency than the other two methods. For example, PEST method significantly decreased RMSE of soil water content, nitrate concentration, crop yield and leaf area index by 61.8%, 23.5%, 73.6% and 23.3%, respectively. Furthermore, the accuracy of simulated water content, nitrate concentration and crop yields were significantly improved by using PEST method. However, there were no significant improvements for the soil nitrogen concentrations and leaf area index, compared to the trial-and-error method. With sensitivity analysis, we identified 18 key parameters that had relatively higher sensitivity. Among these 18 parameters, soil water hydraulic parameters and crop genetic parameters had higher sensitivity than soil nitrogen transformation parameters. Among soil water hydraulic parameters, the soil saturated water content had the highest sensitivity; among crop parameters, the total cumulative available temperature and maximum specific leaf area showed the highest sensitivity; and among soil nitrogen transformation parameters, the maximum soil nitrification rate showed the highest sensitivity. Overall, the sensitivity of nitrogen transformation parameters was generally lower compared with those of soil hydraulic parameters and crop parameters. The sensitivity of crop parameters was significantly different between wheat （C3 crop） and maize （C4 crop）, e.g., the maximum root depth and the maximum assimilation rate for maize showed a higher sensitivity than those of wheat, suggesting that model calibration and validation should be crop specific. The PEST method not only greatly saved time for model calibration, but also achieved significant higher simulation accuracy than that by trial-and-error method. In conclusion, the PEST parameter optimization program is a useful tool and should be adopted in calibration and application of soil-crop models.