Global Sensitivity Analysis of Environmental Models: Convergence, Robustness and Accuracy Analysis

Abstract:

Sensitivity analysis aims to characterize the impact that changes in model input factors (e.g. the parameters) have on the model output (e.g. simulated streamflow). It is a valuable diagnostic tool for model understanding and for model improvement, it enhances calibration efficiency, and it supports uncertainty and scenario analysis. It is of particular interest for environmental models because they are often complex, non-linear, non-monotonic and exhibit strong interactions between their parameters.

However, sensitivity analysis has to be carefully implemented to produce reliable results at moderate computational cost. For example, sample size can have a strong impact on the results and has to be carefully chosen. Yet, there is little guidance available for this step in environmental modelling. The objective of the present study is to provide guidelines for a robust sensitivity analysis, in order to support modellers in making appropriate choices for its implementation and in interpreting its outcome.

We considered hydrological models with increasing level of complexity. We tested four sensitivity analysis methods, Regional Sensitivity Analysis, Method of Morris, a density-based (PAWN) and a variance-based (Sobol) method. The convergence and variability of sensitivity indices were investigated. We used bootstrapping to assess and improve the robustness of sensitivity indices even for limited sample sizes. Finally, we propose a quantitative validation approach for sensitivity analysis based on the Kolmogorov-Smirnov statistics.