Computationally Inexpensive Identification of Non-Informative Model Parameters
Abstract:Sensitivity analysis is used, for example, to identify parameters which induce the largest variability in model output and are thus informative during calibration. Variance-based techniques are employed for this purpose, which unfortunately require a large number of model evaluations and are thus ineligible for complex environmental models.
We developed, therefore, a computational inexpensive screening method, which is based on Elementary Effects, that automatically separates informative and non-informative model parameters.
The method was tested using the mesoscale hydrologic model (mHM) with 52 parameters. The model was applied in three European catchments with different hydrological characteristics, i.e. Neckar (Germany), Sava (Slovenia), and Guadalquivir (Spain). The method identified the same informative parameters as the standard Sobol method but with less than 1% of model runs. In Germany and Slovenia, 22 of 52 parameters were informative mostly in the formulations of evapotranspiration, interflow and percolation. In Spain 19 of 52 parameters were informative with an increased importance of soil parameters.
We showed further that Sobol' indexes calculated for the subset of informative parameters are practically the same as Sobol' indexes before the screening but the number of model runs was reduced by more than 50%.
The model mHM was then calibrated twice in the three test catchments. First all 52 parameters were taken into account and then only the informative parameters were calibrated while all others are kept fixed. The Nash-Sutcliffe efficiencies were 0.87 and 0.83 in Germany, 0.89 and 0.88 in Slovenia, and 0.86 and 0.85 in Spain, respectively. This minor loss of at most 4% in model performance comes along with a substantial decrease of at least 65% in model evaluations.
In summary, we propose an efficient screening method to identify non-informative model parameters that can be discarded during further applications. We have shown that sensitivity analysis keeps its information content even after screening. We could further demonstrate no loss of model performance after screening and subsequent calibration. However, total model evaluations could be reduced by at least 50% in both applications.