Multisite-multivariable sensitivity analysis of distributed watershed models: Enhancing the perceptions from computationally frugal methods

• We examined 36 informal and 1 formal likelihood functions for SA of the SWAT model.
• Formal lik. fcn. replicated results from 36 informal fcns. with less computations.
• There exists a strong correlation between sensitivity measures of nearby stations.
• Dominant H/WQ processes in upstream areas are different from those in downstream.
• Both approaches identified sensitive and insensitive with high confidence.

This paper assesses the impact of different likelihood functions in identifying sensitive parameters of the highly parameterized, spatially distributed Soil and Water Assessment Tool (SWAT) watershed model for multiple variables at multiple sites. The global one-factor-at-a-time (OAT) method of Morris was used for sensitivity analysis of streamflow, combined nitrate (NO3) and nitrite (NO2) fluxes, and total phosphorous (TP) at five gage stations in a primarily agricultural watershed in the Midwestern United States. The Morris method was analyzed for 36 combinations of informal likelihood functions, gage stations, and SWAT model output responses, including relative error mass balance (BIAS), Nash–Sutcliffe efficiency (NSE) coefficient, and root mean square error (RMSE) for peak and low fluxes, and one formal likelihood function that aggregates information content from multiple sites and multiple variables using 65 SWAT parameters. The correlation between sensitivity measures from different likelihood functions was also assessed using the Spearman's rank correlation coefficient. Sensitivity of parameters using different likelihood functions was highly variable, although sensitivity of streamflow and TP showed a high correlation. A stronger correlation between sensitivity of nutrient fluxes at the upstream stations as well as the stations closer to the watershed outlets was evident. Comparison of the combined rank of parameters from informal likelihood functions and the ranks obtained from the formal likelihood function confirmed formal likelihood function ability to effectively identify both sensitive and insensitive parameters with less computational and analysis burden. Uncertainty analysis of the Morris results using bootstrap replications showed that both formal and informal likelihood functions identified sensitive parameters with high confidence.