Improving physically based snow simulations by assimilating snow depths using the particle filter

Data assimilation can help to ensure that model results remain close to observations despite potential errors in the model, parameters, and inputs. In this study, we test whether assimilation of snow depth observations using the particle filter, a generic data assimilation method, improves the results of a multilayer energy-balance snow model, and compare the results against a direct insertion method. At the field site Col de Porte in France, the particle filter reduces errors in SWE, snowpack runoff, and soil temperature when forcing the model with coarse resolution reanalysis data, which is a typical input scenario for operational simulations. For those variables, the model performance after assimilation of snow depths is similar to model performance when forcing with high-quality, locally observed input data. Using the particle filter, we could also estimate a snowfall correction factor accurately at Col de Porte. The assimilation of snow depths also improves forecasts with lead-times of, at least, 7 days. At further 40 sites in Switzerland, the assimilation of snow depths in a model forced with numerical weather prediction data reduces the root-mean-squared-error for SWE by 64% compared to the model without assimilation. The direct insertion method shows similar performance as the particle filter, but is likely to produce inconsistencies between modeled variables. The particle filter, on the other hand, avoids such limitations without loss of performance. The methods proposed in this study efficiently reduces errors in snow simulations, seems applicable for different climatic and geographic regions, and are easy to deploy.