Bias adjustment and skill assessment of sub-seasonal and seasonal snow depth forecasts for Finland

The skilful forecasts of snow depth and the snow cover on the sub-seasonal and seasonal scale would be useful for many sectors. In the EU H2020 e-shape project we developed and piloted sub-seasonal and seasonal forecasts involving the snow. The final products were tailored forecasts of winter street maintenance for the City of Helsinki and forecasts of winter tyre season and safe driving conditions in Finland for the tyre company Vianor. We used the extended-range forecasts of the ECMWF and SEAS5 long-range forecasts of ECMWF available from C3S (at the reduced spatial resolution). For the evaluation, seasonal forecasts from MARS (at the original resolution) were also used. On the seasonal scale, the different spatial resolutions of the CDS and MARS can influence the results of the surface-based variables, such as snow, more than, for example, 2m temperature. 

As snow is somewhat slowly evolving surface phenomenon, its forecasts might remain skilful longer than atmospheric variables, such as 2m temperature. There are also seasonal differences in predictability of snow. This is explained partly by the amount of snow present, but we also examine if the melting snow in spring is more predictable than the transient snow in autumn before the permanent snow cover.  

We tested several bias-adjustment methods, such as the mean bias removal and the parametric quantile mapping (QM), but, based on the preliminary results, we concentrated on non-parametric empirical QM (EQM) and ensemble model output statistics (EMOS). As a non-parametric method, EQM is very flexible and does not require many preconditions on the data, but it does not ensure forecast reliability and coherence and is therefore not a wholly satisfactory method. The distributions of weekly and monthly values are rather Gaussian and encourage using EMOS, but while the non-Gaussian aspects of the weekly and monthly distributions can be mitigated using the censored or truncated Gaussian distributions, the model can break when only small amounts of snow are present. Also, a constant spread should be used in EMOS, as standard deviation of ensemble members is no longer a good estimate of uncertainty for longer time scales.  

We used both in-situ observations and the ERA5 reanalysis for verification and the problem of representativeness of observation for verification was examined. Using point-like in-situ observations is too optimistic for verification of sub-seasonal and seasonal forecasts, and smoothed analysis of ERA5, interpolated to the same resolution as the forecasts, gives more realistic results. However, the interpretation of the ERA5-based results might be a bit challenging for the end-user, who is often interested in the very small-scale features and would like to “zoom in” as much as possible. Here training and co-design with the end-users are needed to fully exploit the value of forecasts.