Name
Evaluating a hierarchy of bias correction methods for reanalysis SWE estimates in northern Canada
Description
Accurate estimation of snow water equivalent (SWE) is essential for the management of water resources in snowmelt-dominated regions like Canada’s north. Melting snow contributes the majority of runoff to rivers and lakes, and it is an important driver of springtime soil moisture. Due to the sparse network of snow-survey sites, climate reanalysis systems are commonly used to provide long-term, spatially consistent SWE estimates. However, even the latest generation reanalysis systems, such as ECMWF’s ERA5, are prone to spatio-temporal biases due to observational errors in data assimilation and physical uncertainties in the reanalysis system itself. Here, we evaluate a hierarchy of linear and non-linear statistical methods for bias-correcting SWE estimates from ERA5-Land over Arctic Canada. The statistical models are trained using predictors like geolocation, elevation, meteorological forcing, and reanalysis SWE to predict in-situ SWE at 163 sites across northern Canada. Our results show that, compared to simpler models like ordinary least-squares regression (OLS), and nonlinear Random Forest (RF) regression performed better in reducing SWE biases both spatially and temporally (RMSE reduction by 64% and 51% over default ERA5-Land by RF and OLS respectively), suggesting that the RF model best captures nonlinear processes contributing to the biases. The bias-corrected SWE dataset will be used to validate and improve the estimation of snowmelt runoff, and to gain an improved understanding of the impact of SWE biases on snowmelt and subsequent spring soil moisture.