Evaluation of the performance of a dataset for snow water equivalent in southern Quebec, obtained using a spatialized particle filter

Abstract. A dataset of daily snow water equivalent (SWE) grids was created for southern Quebec, in eastern Canada, using the HYDROTEL snow model and a spatialized particle filter to assimilate snow data observations. Our aim was to compare this dataset, called "UdeS Grids", to independent observation data used as reference, as well as to five other gridded datasets (ERA5-Land, Crocus-ERA5, MERRA-2, SNODAS, and C3S Satellite SWE). First, we evaluated the ability of these datasets to estimate SWE using correlation coefficients, relative bias, and the normalized root mean square error (nRMSE). Next, we evaluated their ability to represent snow cover duration using three indicators: start date of snow cover, start date of snowmelt, and end date of snow cover. Our findings show that the best performing dataset is UdeS Grids, followed by ERA5-Land and Crocus-ERA5, while datasets derived from satellite data, such as MERRA-2 and C3S Satellite SWE, perform the worst. The UdeS Grids dataset produces the best representation of snow cover duration compared to the reference data. UdeS Grids is well correlated to reference data, and has a relatively low nRMSE. However, it underestimates SWE values, particularly SWE_max. Data from Crocus-ERA5 are of similar magnitude to reference data, but show more rapid snowmelt, resulting in a much shorter duration of melting. ERA5-Land and SNODAS overestimate SWE values, and melting begins two to three weeks later than for other datasets. Furthermore, these two datasets have much longer melting periods than the other datasets. MERRA-2 and C3S Satellite SWE underestimate reference data by more than 50%.