Classification of extra-tropical cyclones with cluster analysis based on measures of intensity

Extra-Tropical Cyclones (ETC) cause most of the variability in weather in Europe and are associated with strong winds, heavy rainfall and storm surges which cause damage to infrastructure and can even lead to human fatalities. Although ETCs vary considerably from one system to another, the classification of ETCs based on their common features is a standard practice in meteorology. Classifications simplify the study of ETCs and thus improve understanding and predictability of them. Due to a large number of factors affecting the development and structure of ETCs, various classifications have been established. ETC classifications are based on e.g. conceptual or idealized models, forcing factors, dynamical and synoptic features, or impacts. In this study we construct a dataset of dynamical and impact-based measures of ETC intensity. The aim is to create a classification of ETCs by using this dataset as input in a cluster analysis.

ETCs and their intensity measures were studied using ERA5 reanalysis data from 1979 to 2022. Only the cold season (October--March) was considered in the analysis since that is when the strongest ETCs occur most often. ETC tracks were produced by using 850-hPa relative vorticity as input to the feature tracking software TRACK. To focus on the most relevant ETCs affecting Europe, only tracks in the North Atlantic and Europe were included and stationary and short-lived systems were excluded. Analysed intensity measures included 850-hPa relative vorticity, Mean Sea Level Pressure (MSLP), 850-hPa, 925-hPa and 10-m wind speed, 10-m wind gust, instantaneous Storm Severity Index (SSI), accumulated SSI and wind footprint. Sparse Principal Component Analysis (PCA) was performed to analyse the dependency of the intensity measures on one another and to select input features for the cluster analysis. The cluster analysis was performed using Gaussian Mixture Modelling (GMM).

Sparse PCA with four principal components indicated that all wind speed measures were grouped together as the first component and 850-hPa wind speed was grouped with 850-hPa relative vorticity as the third component. Wind footprint and MSLP were alone in the second and fourth components, respectively. The SSI measures were not represented, which means they did not explain variance in the dataset but formed a separate group of features. By including 850-hPa relative vorticity, MSLP, 850-hPa wind speed, wind footprint and instantaneous SSI, each group of features was represented in the cluster analysis. The number of clusters in the GMM was increased incrementally by one until each cluster was clearly separate from the others in at least one dimension. With seven clusters, weak ETCs, ETCs with high SSI, and ETCs with similar vorticity, MSLP and wind speed but unequal wind footprint and vice versa, were able to be separated into their own classes and their characteristics investigated.