Inference of snow dynamics from streamflow observations: choose your metrics wisely

High elevation streamflow integrates the hydrological response to snow accumulation and melt. Accordingly, streamflow observations hold valuable but often underutilized information about snow water equivalent (SWE), offering a means to reconstruct historical SWE dynamics. We develop an inverse hydrological modelling framework to derive SWE estimates from streamflow: the framework generates a range of prior SWE scenarios, feeds them into a hydrological model and computes their likelihood based on how well corresponding streamflow simulations match observed streamflow. A critical step hereby is the choice of model performance metrics to be used as likelihood functions. To test our framework, we perform a range of tests in a synthetical setting, where we use known SWE data that we feed into the hydrological model and then apply the inversion framework to retrieve the SWE time series from the streamflow alone. The goal of this synthetic setting is to determine which streamflow metrics select realistic SWE scenarios (measured in terms of errors between the known SWE time series and generated scenarios).

Our results reveal that classical streamflow metrics, such as the Kling-Gupta and Nash-Sutcliffe efficiencies, show no correlation with any SWE timing or magnitude error metrics. Accordingly, minimizing these streamflow metrics does not result in an efficient selection of  SWE scenarios. In contrast, minimizing the mismatch of selected streamflow signatures, such as the baseflow index and the mean melt-season discharge, does lead to a selection of SWE scenarios with smaller errors. Overall, however, our results show that correlations between streamflow performance metrics and SWE performance metrics are generally weak and show significant year-to-year variability, indicating that streamflow metrics are often not informative for reconstructing SWE. 

Our synthetic modeling experiments are conducted in the Dischma catchment in Switzerland, using the OSHD Swiss SWE reanalysis product as the synthetic SWE observations (Mott et al., 2023). Synthetic streamflow observations are generated by feeding OSHD snow melt and MeteoSwiss rainfall into the hydrological model. 

Our findings are relevant for future studies aiming to evaluate or calibrate SWE simulations against streamflow observations, and will help us in the application of the inverse hydrological framework to real-world SWE reconstructions.

 

Mott, R., Winstral, A., Cluzet, B., Helbig, N., Magnusson, J., Mazzotti, G., Quéno, L., Schirmer, M., Webster, C., and Jonas, T.: Operational snow-hydrological modeling for Switzerland, Front. Earth Sci., 11, 1228158, https://doi.org/10.3389/feart.2023.1228158, 2023.