The fifth reanalysis generated by the European Centre for Medium-Range Weather Forecasts (ECMWF), ERA5 , uses the Integrated Forecasting System (IFS, cycle 41r2). In this study, we use ERA5 data from 1979–2020 in the Eastern Mediterranean domain (27–53° N, 7–38° E), interpolated to a horizontal resolution of 0.5°. Specifically, we analyse PV on isentropic levels between 320 and 340 K at 5 K intervals, SLP, precipitation of both convective and large-scale components, 2 m temperature and 10 m winds.

Cold and hot extremes are identified based on the 2 m temperature, averaged daily to eliminate diurnal differences as a factor in temperature variability. To identify the strongest daily deviations across all seasons, we considered at each grid point the daily 2 m temperature anomalies (deviations) from the long-term (1979–2019) monthly mean. A cold or hot daily extreme is identified as a day with a temperature anomaly outside the 5th or 95th percentile range at each grid cell .

We quantify precipitation of two data sources.

Short-term forecasts of precipitation are taken from ERA5, providing 1 h accumulated large-scale and convective precipitation from 12–18 h forecasts. Shorter lead times are avoided due to potential spinup biases.

ERA5 precipitation data were validated against IMERG satellite-based observations for their common period (2000–2020) to ensure their reliability for climatological analyses conducted in this study. This validation allows us to confidently extend our analysis to a longer climatological period (1979–2020) and to analyse a larger sample of cyclones. Additionally, the ERA5 dataset enables differentiation between convective and large-scale precipitation, facilitating a detailed examination of precipitation mechanisms associated with each EMC group. We analyse ERA5 precipitation accumulated for 1 and 3 d, beginning at 06:00 UTC a.m. of the day of minimum cyclone SLP to match the daily accumulated station data.

Considering the variability that exists among cyclone identification and tracking methods, which lead to diverging results in the Mediterranean, here we used a robust cyclone track dataset that takes into account several methods. These composite tracks combine outputs from ten cyclone-tracking approaches from the period 1979–2020, derived from the ERA5 reanalysis hourly intervals . We select composite tracks with a confidence level of 5, indicating that at least 5 cyclone detection methods agree on the cyclone track. This choice is a compromise between too few resulting tracks for a higher confidence level and over-representation of spurious summer cyclones for lower confidence intervals. We then select cyclone tracks that are found within the domain of 27–38° N, 27–38° E (Fig. b) for at least one time step, resulting in 686 cyclone tracks. The peak time of the cyclone is defined based on the minimum SLP along the track within the domain.

We classify EMCs during their peak intensity (minimum SLP) by analysing upper-level isentropic PV fields. We focus on the 320–340 K isentropic surfaces at 5 K intervals, vertically averaged within the fixed extended geographical region of 27–53° N and 7–38° E. This domain is more extended than the domain for the surface cyclone features, to include the extent of all upper-tropospheric large-scale PV features that can affect the cyclone domain. To categorise the PV patterns, we apply a SOM algorithm , an unsupervised neural-network method designed to project high-dimensional data onto a lower-dimensional lattice while preserving topological relationships. The SOM has been widely used in synoptic climatology to identify recurring large-scale atmospheric circulation patterns e.g.,.

Using the SOM algorithm, provided by MathWorks, we classify the PV patterns. Each EMC member is eventually attributed to one of the clusters, based on the similarity of their PV distribution.

Six clusters were selected to represent distinct recurring PV patterns, effectively representing the distinct upper-level atmospheric states. Increasing the number of clusters does not reveal new patterns; instead, it dilutes the significance of our impact analysis. Conversely, selecting a lower number of clusters may enhance the statistical significance of the results, but it fails to capture the full range of upper-level atmospheric states present in this region during EMC events.

The SOM was implemented using a 2×3 rectangular lattice (Fig. ) with a hexagonal topology and Euclidean distance metric. The training consisted of 450 initialisation iterations followed by 550 training epochs. Each EMC is assigned to the SOM node (hereafter referred to as clusters for consistency with Mediterranean synoptic classification studies, ) by matching the EMC's PV structure to the most similar SOM cluster PV structures.

Since the upper-level PV structure associated with EMCs exhibits substantial variability, some EMCs may be assigned to different clusters in different runs of the SOM. This variability can cause individual PV patterns to shift between two clusters across different algorithm iterations. To mitigate this issue and ensure a more robust classification, we ran the SOM algorithm 100 times. The cluster assignment for each EMC was determined based on the most frequently occurring classification across these runs, requiring that the selected cluster appeared in at least 60 runs. This statistical approach reduces sensitivity to minor variations in clustering and provides a more stable and representative classification of the upper-level PV structures. Notably, only 11 EMCs changed their cluster assignments between runs, indicating a high degree of stability in our final classification.

We diagnose the maximum deepening rate of EMCs using units of Bergeron . This rate is normalised by the cyclone latitude and is given at time t by: 1Bergeront=sin⁡(60°)sin⁡(ϕt)SLPt-12h-SLPt12h where ϕt is the latitude of the cyclone center at time t, SLPt-12h is the sea level pressure (SLP) 12 h before time t and SLPt is the sea level pressure (SLP) at the time t.

Different from a commonly used 24 h window for cyclones over oceans, we employ a 12 h window due to the relatively short tracks (31 h on average inside the domain), less than the typical 3–3.5 d for systems forming over oceans. The window is centred around time t and moved along each cyclone track to account for the maximal deepening rate for each track. Based on this criterion, 8 EMCs with deepening rates greater than 1 Bergeron are classified as explosive cyclones.

To evaluate long-term trends of EMC frequency while reducing short-term inter-annual variability, we first smooth the data using a 3 year moving mean. The statistical significance of long-term trends is evaluated with the Mann–Kendall (MK) test for the smoothed data, using a significance level of the test, α=0.1 and resulting in a binary result of 1 indicating a significant trend, or 0 if not.

It is important to note that the trends are calculated based on a relatively small number of cyclones (order of 100). However, despite the small sample size, the trend may indicate the beginning of a pattern which may continue in the warmer future. Therefore, when interpreting the trends, this limitation should be considered, particularly regarding long-term changes and statistical significance.

The analysis uses annual periods running from 1 August–31 July rather than calendar years. Therefore, the first and last seasons in the record (e.g., the 1979 and 2020 seasons) are partially truncated, and the trend calculation does not include the calendar year 2020, as the corresponding season (1 August 2020–31 July 2021) extends beyond the dataset period.

A total of 686 EMCs were identified entering the study domain during the analysis period from 1979–2020. Figure  summarises the key climatological characteristics of these systems. Panel (a) shows the annual frequency of EMCs, revealing substantial inter-annual variability but no significant long-term trend according to the MK test. The climatological mean track density (Fig. b) highlights preferred regions for centres of EMC presence, particularly north of Cyprus and along the Turkish coast. EMC occurrence is enhanced during the winter and early spring months, with peak activity from December to April at 14 % occurrence frequency (Fig. c). These EMC characteristics are in line with known regional patterns , and support the applicability of our chosen EMC track data.

Six meaningful clusters are selected for the SOM analysis, organised into a so-called map that reflects gradual transitions between EMC types, with neighbouring clusters sharing similar PV distributions. Most clusters are characterised by high PV in the upper troposphere, typically organised in a trough-like structure upstream of the EMC centre. Each cluster is characterised by the composite mean PV and SLP of its EMC members (Fig. ). Cluster 1 is marked by high PV values and strongly-tilted wave-breaking structure with a south-west-to-north-east-tilted trough, while the neighbouring Cluster 3 exhibits a similar PV pattern but with weaker wave-breaking and PV values and a less pronounced ridge upstream, over the central Mediterranean and Europe. In contrast, Cluster 2, which includes the largest fraction of EMCs (22 %), has local PV values comparable to values seen in Cluster 1 over the Middle East but is marked by a wide and deep trough and no wave breaking. Cluster 4 shows a more zonal PV distribution with lower PV and PV gradients compared to its neighbouring cluster 2. Clusters 5 and 6 are the least frequent, accounting for 13 % and 10 % of the EMCs, respectively. Both are characterised by relatively low PV in the eastern Mediterranean region. However, Cluster 5 remains associated with wave breaking, while Cluster 6 exhibits enhanced PV in the northwestern part of the domain, north-west of Greece and Italy.

Considering the composite SLP distribution, Cluster 1 shows a local minimum in SLP over Cyprus, indicative of relatively deep cyclones, with a composite mean central pressure of 1008.6 hPa. Cluster 3 is similar but has, on average, a shallower (1009.7 hPa) and more northeastern cyclone centre, located between the coasts of east Cyprus and west Lebanon, with a markedly less pronounced upstream ridge. Cluster 2 has the deepest SLP minima (1007.2 hPa), located west of Cyprus, while Cluster 4, with a shallower intensity (1008 hPa), still exhibits a distinct low in the climatological cyclogenesis region near Cyprus. Differently, Clusters 5 and 6 show SLP minima in the southeast Mediterranean near the Red Sea, although without a clear closed low within the domain, suggesting more variability in the location of the cyclone centres. Indeed, Fig.  shows that the location of the minimum SLP of individual tracks has a larger spread in Clusters 5 and 6. The 8 explosive EMCs, highlighted in Fig. , can be found in Clusters 1 and 2, indicating a preference for PV structures associated with wintertime cyclone configurations, with one additional exceptional case identified in Cluster 5. In fact, most of the cyclones (55 %) are already in their weakening phase when they enter the domain, as also indicated by the clustering of minimum SLP locations near the 27 E meridian. This suggests that for a significant number of systems, the mature phase of the cyclones occurs upstream of the domain boundary, and they affect the eastern Mediterranean domain at their decay phase. The results also highlight how surface SLP features – such as the location and depth of minima – correspond closely to upper-level PV structures, particularly the positioning of the PV troughs and ridges that guide cyclogenesis in this region.

While regional EMCs occur mostly during the extended winter season (Fig. ), the SOM classification highlights a more refined and distinguished seasonal signature for each of the resulting clusters. In January and December, there is a higher frequency of Clusters 1 and 3 (Fig. ), indicating that these atmospheric patterns are more prevalent at the onset of winter. This is consistent with the typical intensification of upper-level troughs and stronger baroclinic activity during winter, favouring the development of pronounced PV anomalies. In contrast, Clusters 2 and 4 primarily comprise the late winter, with Cluster 2 occurring mostly between January to March and early spring, while Cluster 4 occurs most frequently during March and April. This suggests that there is a shift in the regional dynamical cyclone lifecycles as winter progresses into spring. As expected by their overall lower PV, Clusters 5 and 6 appear mostly during the transitional seasons of autumn and spring (respectively). These clusters appear even in summer, but with lower frequencies, highlighting their association with different atmospheric conditions typical of these periods. These findings highlight the seasonal variability of the EMCs PV patterns and underlying dynamical changes, with specific clusters associated with distinct phases of the seasonal cycle. At the same time, these results suggest that a seasonal distinction alone would not resolve the full variability of EMCs, as most clusters also occur outside of their typical season, and different EMC types occur at any given month.

To further quantify the dynamical characteristics of each cluster, Table  summarises mean and standard deviation values of key EMC track properties, including deepening rate, mean minimum SLP, duration within the domain, and distance travelled both within the defined domain and along the entire cyclone track. Cluster 1 is characterised by a relatively high deepening rate and the longest duration (∼41.84h), along with moderate distances travelled both within the domain (∼978km) and overall (∼1402km). This indicates long-lived, dynamically active systems that tend to remain partly stationary during their time in the research domain, particularly over the Cyprus cyclogenesis region, most likely benefiting from persistent baroclinic conditions. In contrast, Cluster 2 exhibits a lower deepening rate and shorter duration (∼29.46h) but travels significant distances within (∼927km) and outside the domain (∼1794km). EMCs in Cluster 3 display the highest deepening rate, although this cluster has no explosive cyclones. They have a moderate duration (∼33.44h) and moderate total distance (∼1404km) but the longest distance within the domain (∼1022km). Cluster 4 contains the shortest-lived (∼21.39h) and weakest cyclones (lowest deepening rate), travelling shorter distances inside the domain (∼677km) yet covering the largest total distances (∼1852km). This indicates transient systems that quickly pass through the domain without significant intensification but continue travelling extensively to the East of the domain. Clusters 5 and 6, typical of transition seasons and summer, reflect characteristics of thermally driven or weakly baroclinic cyclones like Sharav lows, as will be shown in Sect. . The EMCs in Cluster 5 show moderate deepening rates and duration (∼33.98h), with moderate distances travelled inside (∼911km) but the shortest overall distances (∼1312km), consistent with relatively stationary cyclones influenced by weak upper-level forcing and local thermal gradients. Conversely, EMCs in Cluster 6 exhibit the lowest deepening rate and short durations (∼23.64h) but relatively large overall distances (∼1722km). These features suggest weaker systems, likely influenced by transient upper-level disturbances or weaker baroclinicity, yet displaying notable mobility beyond the study domain. These characteristics align well with the seasonal variability observed in EMC behaviour in the region.

It is important to consider the distribution of the accumulated precipitation that constructs the spatial composite mean patterns. While the mean provides a view of most cases, it may obscure information about extremes, their number and location. Moreover, relying solely on the mean values does not indicate whether these values stem from a few intense events or if the precipitation distribution is more evenly spread across the cases. Therefore, for each of the 686 EMCs we examined the distribution of ERA5 3 d accumulated precipitation within the research domain and cluster type (Fig. a). Clusters 1, 2, 3 and 4 are characterised by a relatively uniform distribution, indicating that precipitation events within these clusters consistently cover a wide range of intensities, and there are no isolated heavy rainfall events that dominate the distribution. Conversely, other clusters are primarily influenced by a few individual events contributing substantial rainfall amounts. Cluster 5, for instance, includes many events with little or no precipitation and a long tail of extremes, while Cluster 6 features consistently low precipitation values throughout. Notably, many of the extreme precipitation values in the distributions are associated with explosive EMCs, which are marked by the letter “b” in Fig. a.

To better capture localised extremes, we complement this analysis with a grid-point-based approach using 24 h accumulated ERA5 precipitation. For each EMC, we count the number of grid points exceeding different precipitation thresholds within the domain and construct distributions of grid-point precipitation intensities (Fig. b). This shorter accumulation period is more representative of local extreme events and reduces the smoothing effect inherent in longer temporal and spatial averaging.

The results show that cluster 1 EMCs host at least a third of localised extremes, followed by cluster 2 and 3 EMCs, despite the latter two occurring more often (over 20 % of the time compared to 16 % in cluster 1). While all clusters exhibit a decrease in the number of grid points with increasing precipitation intensity, Cluster 5 is characterised by a relatively higher fraction of high-intensity grid-point precipitation compared to the other clusters. This indicates that despite its generally low domain-averaged precipitation, Cluster 5 is more strongly associated with localised extreme rainfall events consistent with its summer-autumn seasonality.

Overall, the combined analyses demonstrate that spatial averages can mask important differences in the structure of precipitation extremes, and highlight the need to consider both domain-averaged and grid-point-based metrics when assessing the impact of EMCs on regional precipitation.

To further understand surface impacts of EMCs, we examine the temperature anomalies associated with the different PV patterns. Figure  presents the composite mean 2 m temperature anomalies calculated from the monthly mean climatology for each cluster. Overlaid are the grid points where more than 15 % of the EMCs experienced extreme cold or warm daily temperature anomalies (below the 5th or above the 95th extreme percentile of the distribution, respectively). As either cold or warm anomalies are expected with a potential frontal passage, composites are shown for the time of minimum SLP and for 12 h after.

At the time of minimum SLP, all clusters except for Cluster 3 exhibit positive anomalies in the eastern part of the domain, with a strong signature over land areas, concurrent with the warm sector of the cyclone. Clusters 4 and 6, in particular, demonstrate extremely hot temperatures in the southern part of the domain. These hot extremes are known as Sharav heat lows in the region, conductive for warm, dry, and often dusty extremes in spring . At the same time, over the sea, all clusters but Cluster 6 exhibit cold anomalies. Winter Clusters 1 and 2 exhibit strong cold anomalies (compared to their respective months). In these clusters, extremely cold temperatures are seen to the west and south of the cyclone centre.

Upon the passage of the cyclones, 12 h after peak intensity, most of the domain becomes colder (Fig. ). In Clusters 1 and 2, the warm sector associated with the cyclones traverses the domain, leading to anomalously cold temperatures throughout the region. Cluster 3, which did not exhibit any warm anomalies, shows the domain becoming even colder (at t=+12), especially over land; this may be attributed to the upper-level PV pattern indicating strong air advection from the north. For Clusters 4–6, while the warm anomalies over land persist, their intensity diminishes. Notably, Clusters 4 and 6 maintain pronounced warm extremes over land. This analysis highlights the frontal behaviour of cyclones associated with winter seasons and the presence of a warm core in cyclones more typical of late spring and summer, particularly evident in Clusters 4 and 6. This cluster-specific analysis highlights the strong connection between upper-level PV patterns and surface temperature anomalies and extremes.

It should be noted, however, that the timing of the minimum SLP may influence the results. While the distribution of the time of the minimum SLP is generally uniform, certain times are somewhat favoured and could therefore affect the composite mean, particularly when comparing differences between two specific time steps. To assess the robustness of our findings, we also examined temperature differences using a 24 h interval (not shown). Although the magnitude of the anomalies is weaker, the overall patterns and behaviour remained consistent.

The results underscore that, on average, EMCs introduce temperature anomalies and often extremes to the region. This finding emphasises the importance of understanding the link between the upper-level atmospheric state and its impact near the surface, which could not be implied by the SLP pattern alone.

To better connect the large-scale circulation patterns with their surface impacts, the spatial distribution of 10 m wind speed anomalies is analysed for each cluster. The 10 m wind speed anomalies at the time of minimum SLP, shown in Fig. , reveal clear spatial differences in both magnitude and flow structure among the clusters. In the winter clusters (Clusters 1–3), extensive regions with wind speeds exceeding an anomaly of up to 7–8 ms-1 appear over the eastern Mediterranean Sea. The strongest anomalies are concentrated along the coasts of Egypt, Israel, and Cyprus, and inland over Jordan and northwestern Saudi Arabia.

In Cluster 1, the maximum anomaly is located north of Egypt, especially along the western edge, and to the east and west of Cyprus. The wind speed anomaly vectors, which indicate strong winds, show a coherent southwesterly-to-northwesterly flow that curves cyclonically toward the north and northeast. This curvature, together with the intensity of the anomalies, is consistent with circulation around a deep low-pressure system. Cluster 2 shows a similar pattern, with the strongest winds extending from northern Egypt and the southeastern Mediterranean toward the Israeli coast. Wind speeds are high, and the wind direction remains predominantly southwesterly, turning cyclonically toward the Levant. In Cluster 3, the area of enhanced winds is predominantly located south of Turkey. The wind speed anomalies are well aligned and display clear cyclonic curvature, with inflow from the northwest. This pattern is consistent with the upper-level PV structure, which suggests northerly flow, which then turns westerly and subsequently southerly along the Israeli, Lebanese, and Syrian coasts Cluster 4 represents a transition, with weaker and more localised positive anomalies. Wind speed anomalies are clearly weaker than in Clusters 1–3, and the flow remains cyclonic.

Clusters 5 and 6 show markedly weaker anomalies and more heterogeneous patterns of 10 m wind speed anomalies. Wind speed anomalies are mostly below 4 ms-1 across the basin. In Cluster 5, the flow is less organised, with smaller wind magnitudes and more variable directions, although a weak cyclonic circulation still seems to be present.

A consistent feature across clusters 4-6 is the southeastern part of the domain, particularly over the Sinai Peninsula and the northern Red Sea (around 27–29° N, 33–36° E), where wind speeds are significantly lower and associated with negative anomalies. However, these are not statistically significant.

Overall, the strongest and most organised wind patterns in Clusters 1–3 coincide with the regions of deepest SLP in the composites. This suggests that the enhanced cyclonic flow is driven by strong pressure gradients associated with deep cyclones. In contrast, the weaker and less coherent wind patterns in Clusters 4–6 are consistent with shallower SLP fields.

In this section, we examine the inter-annual frequency of EMCs in each cluster. As different types of EMCs have varying impacts on the surface, their respective trend may suggest expected surface changes in the region in the future climate. When considering all clusters together, there is no significant trend in EMC frequency (Fig. ). However, separately examining each cluster reveals that some of the clusters exhibit significant and opposing trends in the observed period (Fig. ). Clusters 5 and 6 demonstrate a significant increase, while a decrease is seen in Cluster 4. Other clusters do not display significant trends. The significance of the trend for Cluster 5 is the strongest, increasing during the period from 1 to 3 EMCs per year on average.

The compensating trends of the different clusters are especially meaningful when taking into account the surface impacts of each EMC type. As Cluster 5 is associated with low precipitation and warm anomalies with sporadic precipitation extremes, a potential shift toward drier and warmer conditions under EMCs is likely to occur if this trend continues. Thus, considering cyclone frequency as a whole is not a sufficient proxy for precipitation trends, and cluster-specific information is key for understanding long-term changes in precipitation patterns, affecting water resources, agriculture, and the health of the ecosystem in the region.