On the accuracy requirement of surrogate models for adequate global sensitivity analysis of urban low-impact development model

Global sensitivity analysis (GSA) is crucial for understanding, simplifying, and applying high fidelity process-based (Hifi) hydrological models. Its application has been hindered by the extensive model evaluations needed for convergence and the associated computational cost. Surrogate models (SMs) can significantly reduce the computational burden of GSA, but its approximation errors can introduce usually uninvestigated GSA errors. We address this gap by investigating SMs-induced GSA error in parameter screening, sensitivity ranking, and sensitivity index valuation. By comparing the converged GSA results from a support vector regression surrogate model (SVR-SM) and the storm water management model (SWMM) in simulating the hydrological response of a small urban watershed to changes in low-impact development (LID) parameters, this study finds that SMs-induced GSA errors increase with higher SMs approximation errors. The relationship between SMs-induced GSA error and SMs approximation error (measured by R²) is consistent across various flow metrics and rainfall intensities. SMs can adequately reproduce the converged GSA results of a Hifi SWMM with only one-thousandth of the original computation time. However, SMs-induced GSA errors may become unacceptable if the R² of SVR-SMs is below 0.96. Our findings highlight the importance of surrogate models’ accuracy in GSA and provide valuable guidance for future GSA applications.