Sensitivity and uncertainty analysis for the DeNitrification–DeComposition model, a case study of modeling soil organic carbon dynamics at a long-term observation site with a rice–bean rotation

Abstract

Process-based modeling is a powerful tool for identifying predominant factors that influence soil organic carbon (SOC) so that rational management decisions can be made for improving soil fertility and increasing carbon sequestration in agricultural soils. To achieve these goals, a sensitivity and uncertainty analysis of modeling output is required. Based on the data collected at a long-term observation site with a 21-year rice–bean planting history, two global sensitivity analysis (SA) methods and six uncertainty analysis (UA) methods were applied and compared for the DeNitrification–DeComposition (DNDC) model to simulate inter-annual SOC changes (dSOC). Morris and Sobol’ SA methods were used to identify the most influential parameters, and then the dSOC uncertainty intervals estimated by Morris, Sobol’, Morris and a joint use of Monte Carlo (MC) with the five most sensitive parameters considered (Morris-MC-5Fac), Morris and a joint use of MC with the four most sensitive parameters considered (Morris-MC-4Fac), DNDC embedded Most Sensitive Factor method (DNDC-MSF), and the user-modified Most Sensitive Factor method with the five most sensitive parameters considered (MSF-user-5Fac) were compared. Sensitivity analysis results indicated that the Morris and Sobol’ methods produced similar sensitivity indices for 11 DNDC model input parameters. The initial SOC, bulk density, amount of manure applied, ratio of crop residue incorporated into soils, and amount of chemical fertilizer applied are the five most influential parameters on dSOC uncertainty at the site. In our case, Morris is recommended for a pure sensitivity analysis due to its low model running cost compared to that of Sobol’. Among the six uncertainty analysis methods, the Sobol’ estimated uncertainty result is considered the most reliable, followed by that estimated with Morris-MC-5Fac, MSF-user-5Fac, Morris-MC-4Fac, Morris, and DNDC-MSF. DNDC-MSF has the lowest model running cost, followed by Morris, MSF-user-5Fac, Sobol’, Morris-MC-4Fac, and Morris-MC-5Fac. Considering both the reliability of the uncertainty analysis results and the model running costs, MSF-user-5Fac is the most efficient. However, it should be noted that MSF-user-5Fac is not suitable for models in which the input parameters and modeling outputs are not monotonically related. For the latter models, Sobol’ is a better choice. The information in this paper benefits DNDC users by providing guidance in proper sensitivity and uncertainty analysis method selection.