Snow Matters: Emulator-Driven Calibration Across a Large Sample of Snow Stations

Calibrating complex physically based hydrological models remains a major challenge due to the high computational demands of traditional optimization algorithms. Furthermore, if conducted, local calibration at several sites can lead to an uneven spatial distribution of parameters, imposing additional challenges when transferring parameter values from gauged to ungauged areas. As a result, simulations from complex models often rely on default parameter values that yield poor model performance. Recent studies have shown that machine learning emulators can speed up calibration and regionalization of model parameters while maintaining predictive accuracy similar to that of traditional optimization algorithms and improving the spatial distribution of parameters. However, most studies using emulators focus on streamflow, while there is a great opportunity to support improved process modelling using large datasets. Here, we focus on snow, a critical component of hydrological systems, and show how improved calibration of snow-related parameters could enhance the consistency of hydrologic model simulations. In this study, we assess whether emulators can (1) improve snow simulations across North America and (2) regionalize snow parameters across the continent. To this end, we use 770 snow stations located in Canada and the United States. We compare the performance of a conceptual model (FUSE: Framework for Understanding Structural Errors) and a physically based model (SUMMA: Structure for Unifying Multiple Modeling Alternatives), each calibrated using both traditional algorithms and emulator-based approaches. Our results show that snow simulations using SUMMA achieve performance comparable to that of FUSE with a fraction of the simulation runs usually required by optimization algorithms, suggesting that complex models can perform similarly to (calibrated) conceptual models. Further, when conducting local calibration, the use of large-sample emulators improves the smoothness of the spatial distribution of parameters, which, for parameter regionalization purposes, translates into smoother spatial distributions of parameter values in large geographic areas. This suggests that emulators can mitigate the effect of the highly irregular response surface during parameter calibration, thereby enhancing the robustness of simulations across large domains. Thus, this work offers new insights into the potential of emulators to enhance process-based modeling and snow representation across large, diverse regions.