On the use of Random Forest as a Global Sensitivity Analysis method: a large-sample application across Europe

Global sensitivity analysis (GSA) is a crucial tool for identifying influential parameters in hydrologic models. GSA methods can guide the selection of calibration parameters, thereby reducing the dimensionality of the parameter space by discarding less influential parameters. This study will explore the use of Random Forest as a GSA method. In particular, feature importance in Random Forest can be interpreted as sensitivity measures of input parameters once the regression task is performed with respect to the output variable of interest. This work will be focused on a large-sample application of the Variable Infiltration Capacity (VIC) model across Europe. A substantial number of Monte Carlo simulations will be carried out for each catchment in order to explore the parameter space and generate a large dataset for the Random Forest regression. Parameters sensitivities will be quantified for the Kling-Gupta Efficiency (KGE) of daily streamflow based on feature importance, and results will be compared against traditional GSA techniques such as the Standardized Regression Coefficients (SRC) and the Regional Sensitivity Analysis (RSA) methods.

Acknowledgments: This study has been funded by a Humboldt Research Fellowship for Postdoctoral Researchers from the Alexander von Humboldt Foundation.