Evaluation and machine-learning-based downscaling of ERA5 snow depth in Central Europe

Reanalyses are currently a major source of information about the state of the troposphere in many meteorological and climatological investigations. However, since they are a product of numerical modeling, they contain some systematic errors, which occur mainly due to the discretisation of a domain, parametrisation of physical processes and numerical approximations. The errors are particularly apparent for meteorological variables which cumulate over time and have great spatial heterogeneity. One example of such a variable is snow depth. In this work, snow depth from ERA5 and ERA5-Land reanalyses was evaluated against observations from over 300 stations in Poland, Czech Republic and Slovakia in winter seasons from 2001 to 2021. Additionally, an attempt was made to reduce existing biases and produce the field in finer resolution based on terrain characteristics derived from a digital elevation model. Verification results show that the difference between modelled and real elevation, together with station rank, are major factors contributing to systematic errors. In complex terrain, bias occurs nearly at every station, while in the lowlands, it is roughly neutral at synoptic stations as their data are assimilated during the production of the reanalyses. For lower-ranked stations, errors are apparently higher. On average, ERA5-Land has lower Root Mean Square Error (RMSE) than ERA5, however, there are stations with greater positive and negative bias. Despite higher resolution, ERA5-Land performs worse than ERA5 for around 30% of stations. Machine learning methods are capable of reducing systematic errors existing in the reanalyses. The greatest improvement occurs especially for sites in complex terrain. Statistical downscaling provides some useful insight in spatial distribution of snow depth, however, its physical consistency needs to be enhanced, e.g., by using physically-constrained machine learning.