Deep learning-based classification of atmospheric circulation types over Europe in a CMIP6 Large Ensemble

The 29 circulation types by Hess & Brezowsky, called “Großwetterlagen”, are one of the most established classification schemes of the large-scale atmospheric circulation influencing Europe. They are widely used in order to assess linkages between atmospheric forcing and surface conditions e.g. extreme events like floods or heat waves. Because of the connection between driving circulation type and extreme event, it is of high interest to understand future changes in the occurrence of circulation types in the context of climate change. Even though the “Großwetterlagen” have been commonly used in conjunction with historical data, only very few studies examine future trends in the frequency distribution of these circulation types using climate models. Potential reasons for this, are the lack of an open-source classification method of the “Großwetterlagen” and their high range of internal variability. Due to the dynamic nature of the large-scale atmospheric circulation in the mid-latitudes, it is highly relevant to consider the range of internal variability when studying future changes in circulation patterns and to separate the climate change signal from noise.

We have therefore developed an open-source, automated method for classifying the “Großwetterlagen” using deep learning. We apply this method to the SMHI-LE, an initial-condition single-model large ensemble of the CMIP6 generation with 50 members on a daily resolution. A convolutional neural network has been trained to classify the circulation patterns using the atmospheric variables sea level pressure and geopotential height at 500 hPa at 5° resolution. The convolutional neural network is trained for this supervised classification task with a long-term historical record of the “Großwetterlagen” covering the 20^th century. It is derived from a subjective catalog of the German Weather Service with daily class affiliations and atmospheric variables from ECMWFs’ reanalysis dataset of the 20th century, ERA-20C.

We present the challenges of the deep learning-based classification of subjectively defined circulation types and quantify the uncertainty range intrinsic to deep neural networks using deep ensembles. We furthermore demonstrate the benefits of this automated classification of “Großwetterlagen” with respect to the application to large datasets of climate model ensembles. Our results show the ensemble-averaged future trends in the occurrence of “Großwetterlagen” and the range of internal variability, including the signal-to-noise ratio, for the CMIP6 SMHI-LE under the SSP37.0 scenario.