and others. By identifying the inputs and parameters that can be targeted to effectively reduce the predictive uncertainty in process-based models with varying complexity (Clark et al., 2015), GSA results can be leveraged to improve model calibration efficiency and accuracy by focusing on parameters that are more sensitive to the observed system responses.For most GSA applications in Earth science domains, including watershed science, ensemble simulations are used to propagate uncertainty from model inputs and parameters to outputs, because oftentimes no explicit functions can be derived to describe their relations (Borgonovo & Plischke, 2016;Pianosi et al., 2016). For multivariate sensitivity analyses of increasingly complex models, a large ensemble of simulations (e.g., 10 3 ) is often required to achieve robust results because the required ensemble size nonlinearly grows with the dimension of parameters and inputs. For example, thousands of ensemble runs are suggested for GSA on multiple parameters (Saltelli et al., 2008) using the popular Sobol method (Sobol, 2001). Nevertheless, despite rapid advances in computing power (Bourzac, 2017), it still is difficult, if not impossible, to afford the thousands of simulations of fully integrated, multiphysics models, such as the Advanced Terrestrial Simulator (