A novel machine learning based bias correctionmethod and its application to sea level in an ensemble of downscaled climate projections

A new machine learning based bias correction method is presented and applied to sea level in a regional climate model for the Baltic and the North Sea. The bias corrections introduced by the method depend on the state of the model it corrects. This contrasts with conventional bias correction methods that operate on distributions of output variables. That is, while conventional correction methods adjusts all modelled sea levels of the same height by the same amount, this method instead adjusts all sea level that occur under the same meteorological conditions by the same amount. Model state dependent corrections allow for better performance on classical skill scores, like correction coefficients, but it also limits the applicability of the method to models that can perform hindcasts. This constrain occurs because the method requires observations and model data from an overlapping time period.

The bias correction method is applied to a large ensemble of dynamically downscaled climate scenario data encompassing many different driving global climate models and representative concentration pathways. The prevalence of significant trends in yearly sea level maximum is found to be independent of emission scenario in our ensemble. This suggests that anthropogenic climate change is not a strong driver of storm surge variability in the area. Moreover, it also suggests that very long datasets of corrected sea levels can be created by merging data from different emission scenarios. A dataset is thus produced that contains over 2600 model years and exists for seven different tide-gauge stations on the Swedish Baltic Sea coast. This dataset is used to estimate return levels for very long return periods by fitting generalized extreme value distributions to block maxima sea level time series. At some stations it is found that the block length used in the return level computation affect the result. This suggests that the commonly used annual maximum approach (i.e. having a block length of one year) is not always applicable for determining return levels for sea level in the area.