A robustness comparison of sampling methods for Sobol’ sensitivity analysis of rainfall-runoff models

Abstract:

Optimal usage of available budget, proper selection of quantities of interest, and appropriate reliability and accuracy help to ensure good practice and outcomes for environmental modeling exercises. Most existing hydrologic models are usually complex in their parameterization and can be computationally expensive, thereby constraining exploration of model properties and behavior. Sensitivity analysis (SA) can be used as a diagnostic tool to identify non-influential model parameters in order to reduce the dimensions of target models, and this can substantially improve the computational efficiency of model calculations such as for calibration and uncertainty quantification purposes.

Model-independent variance-based SA methods have become popular SA methods, among which is the Sobol’ method and its variance-based indices of sensitivity. However, previous studies based on simple test functions have identified that the Sobol’ method coupled with several randomized sampling schemes may provide non-robust sensitivity indices even at reasonably large sample size. In this study, we will compare various sampling methods with randomization methods for use with the Sobol’ SA method. These comparisons will be for several hydrologic models, including the IHACRES rainfall-runoff model and Sacramento soil moisture accounting model to study the robustness of ensuing sensitivity measures.