Application of K-means classification for extraction of atmospheric synoptic conditions leading to the high-frequency sea level extremes of the Adriatic Sea

Sea-level extremes represent a great danger to coastal infrastructure and a daily threat to people living near the coast. These extremes are predicted to become more frequent in the coming years and decades, mostly due to mean sea-level rise. Knowledge of the underlying principles that drive these events is, thus, of greater importance than ever. Our analysis focuses on events of high-frequency sea-level extremes (extremes at periods shorter than 2 hours). Extreme events were extracted from sea level data series measured at six Adriatic Sea tide gauge records. Series lengths were from 16 to 17.5 years. The sea-level data series were split into a training set and a testing set. Splitting was done so that approximately 80% of the series were used for the training and remaining 20% for the testing. K-means classification was then used to associate extremes events of the training period with atmospheric synoptic conditions, represented with the synoptic variables downloaded from the ERA5 reanalysis. The atmospheric variables considered were the ones found by earlier research to be the most important when it comes to generation of intense high-frequency sea-level oscillations. These variables are: (i) temperature at 850 hPa, (ii) mean sea-level pressure and wind at 10 m and (iii) geopotential at 500 hPa. K-means classification was used to find prevailing clusters related to extremes at each of the six tide gauges. After that, the same synoptic variables were downloaded for each day of the testing period. To each tide gauges, and to each day of the testing period, a cluster, previously defined for the training period, was assigned. The idea was to check whether days of known extremes will be correctly clustered. The goodness of the approximations was determined by estimating the distance of the synoptic maps from the clusters. The results show that the testing period days with extremes have a smaller distance from the clusters than random days indicating that there is a potential for prediction of these events.