Comparison of high-resolution climate reanalysis datasets for hydro-climatic impact studies

Abstract. Continuous high-quality meteorological information is needed to describe and understand extreme hydro-climatic events, such as droughts and floods. Information of highest quality relying on observations is often only available on a national level and for few meteorological variables. As an alternative, large-scale climate reanalysis datasets blending model simulations with observations are often used. However, their performance can be biased due to coarse spatial resolution, model uncertainty, and data assimilation biases. Previous studies on the performance of reanalysis datasets either focused on the global scale, on single variables, or on few aspects of the hydro-climate. Therefore, we here conduct a comprehensive spatio-temporal evaluation of different precipitation, temperature and snowfall metrics for four state-of-the-art reanalysis datasets (ERA5, ERA5-Land, CERRA-Land, and CHELSA-v2.1) over complex terrain. We consider climatologies of mean and extreme climate metrics, daily to inter-annual variability, as well as the consistency in long term trends. Further, we compare the representation of extreme events, namely the intensity and severity of the 2003 and 2018 droughts, as well as the 1999 and 2005 floods in Switzerland. The datasets generally show a satisfactory performance for most of these characteristics, exceptions being the representation of snowfall (solid precipitation) and the number of wet days in ERA5 and ERA5-Land. Our results show clear differences in the representation of precipitation among datasets and a substantial improvement of the representation of precipitation in CERRA-Land compared to the other datasets. In contrast to precipitation, temperature is more comparable among datasets, with CERRA-Land and CHELSA showing smaller biases yet a clear increase of bias with elevation. All datasets are able to identify the 2003 and 2018 drought events, however, ERA5, ERA5-Land, and CHELSA overestimate their intensity and severity, while CERRA-Land underestimates it. The 1999 and 2005 floods are overall well represented by all datasets, with CERRA-Land showing the best agreement with observations and the other datasets overestimating the spatial extent of the events. We conclude that overall, CERRA-Land is the most reliable dataset and suitable for a broad range of analyses, particularly for regions where snow processes are relevant and for applications where the representation of daily to inter-annual precipitation variability is important.