基于VIP模型评估不同全局敏感性分析方法有效性及效率

参数敏感性分析可提高复杂过程模型参数优化及模型应用效率。为评估不同全局敏感性分析方法在筛选农田生态系统模型敏感参数时的有效性及其效率，该研究以VIP (Vegetation Interface Processes) 模型模拟华北平原土壤硝态氮为例，对比分析PSUADE (Problem Solving Environment for Uncertainty Analysis and Design Exploration) 提供的8种敏感性分析方法在筛选华北平原农田土壤硝态氮敏感参数时的有效性及其效率。结果表明：在验证敏感性分析方法有效性时，Spearman秩相关系数 (Spearman's correlation coefficient，SPEA) 法和Gaussian Process (GP) 法与其他方法的敏感参数筛选的结果差异较大；多元自适应回归样条 (Multivariate Adaptive Regression Splines，MARS)、Delta Test (DT)、Sum of Trees （SOT）法、McKay法、Morris法和Sobol'法能有效筛选出敏感参数。在分析敏感性方法效率时发现，正交阵列 （Orthogonal Array，OA)和基于拉丁超立方的正交阵列 (Orthogonal Array-Based Latin Hypercubes，OALH) 抽样方式最适于多元自适应回归样条法 (Multivariate Adaptive Regression Splines，MARS)，所需样本量为361。蒙特卡罗 (Monte Carlo，MC) 抽样方式最适合DT和SOT敏感性分析方法，所需最小样本量分别为425和510；与其他抽样方式相比，OALH抽样更适合McKay敏感性分析方法，需要样本量最少；Morris法和Sobol'法所需最小样本量分别为340和810。对于复杂过程模型，可先选用定性敏感性分析方法以较低的计算成本选择敏感的参数，再进行模型参数优化和不确定性分析。

Effectiveness and efficiency of different global sensitivity analysis methods based on the VIP model

Abstract: Global sensitivity analysis generally refers to a sort of numerical approach for the evaluation of the variation in the output of a model, as the input parameter varies. In ecosystem model based on dynamic processes, the parameter verification can be difficult in the model application, due to a complex structure, multiple input parameters, and strong spatial variability. In this case, the parameter sensitivity analysis can effectively identify the factors of main effect, and thereby to improve significantly optimization of parameters and models. Given a fixed value to an insensitive parameter, it is expected to enhance the predict accuracy, while improve computing efficiency of model calibration, verification, and simplification. In this study, a Vegetation Interface Processes (VIP) model was used to simulate the nitrate content of soil in the North China Plain, in order to evaluate the efficiency of different global sensitivity analysis methods in screening sensitive parameters of agro-ecosystem models. Eight sensitivity analysis methods were selected, including six qualitative and two quantitative sensitivity analysis methods, provided by Problem Solving environment for Uncertainty Analysis and Design Exploration (PSUADE). The results show that: 1) The screening information of sensitivity parameter differed greatly in two method of Spearman''s correlation coefficient (SPEA) and the Gaussian Process (GP) from others, when verifying the effectiveness of six qualitative sensitivity analysis methods. The SPEA method cannot effectively identify the decomposition rate of microbial pool, whereas, the GP method cannot identify the decomposition rate of structural litter pool. The similar sensitive parameter screening can be found in the Morris method, multiple adaptive regression spline (MARS), Delta Test (DT), and Sum of Tree (SOT) methods; 2) Sensitive parameters of soil nitrogen cycle that screened by DT, MARS, Morris, and SOT methods were: the potential nitrification rate, urea hydrolysis rate, Michaelis constant, microbial nitrogen-carbon ratio, slow humus nitrogen-carbon ratio, decomposition rate of microbial pool, decomposition rate of structural litter pool, and decomposition rate of metabolic litter pool. In contrast, the SPEA and GP screening showed that the insensitive parameters were the potential nitrification rate, and urea hydrolysis rate, indicating inconsistence with other qualitative sensitivity analysis; 3) When analyzing the efficiency of sensitivity methods, it was found that the sampling methods in the Orthogonal Array (OA) and Orthogonal Array based on Latin Hypercube (OALH) were suitable for Multivariate Adaptive Regression Splines (MARS), indicating that the required sample size was 361. In the DT and SOT methods, the Monte Carlo (MC) was the most suitable for the DT and SOT sensitivity analysis, where the minimum sample sizes were required 425 and 510, respectively. Compared with MARS and DT sensitivity analysis methods, the SOT required larger sample size in the process of screening sensitive parameters. The OALH sampling was the most suitable for the McKay sensitivity analysis methods, while requiring the least sample points. In addition, the minimum sample size can be achieved in the Morris and Sobol'' method with 340 and 810, respectively. In summary, the qualitative sensitivity analysis method required less sample size than the quantitative sensitivity method, but it cannot be used to quantitatively describe the sensitivity of parameters during screening. Therefore, a recommendation during this time can be made that, in the complex system models with many parameters, the qualitative sensitivity analysis method can be used first to screen the preliminary sensitivity parameters at a low computational cost, and then the quantitative analysis of selected sensitive parameters by quantitative sensitivity analysis method.