The observational data set provided by the German Weather Service (DWD) includes hourly surface observations over land from 2001 to mid-2020 containing five parameters: mean sea level pressure (p), 2 m air temperature (T), wind speed at 10 m (v), wind direction at 10 m (d) and precipitation amount (RR). Additionally, the potential temperature (θ) is computed, and θ and v are normalised by their median and 98th percentile respectively, to take the diurnal and seasonal cycles as well as location-specific characteristics into account (θ̃ and ṽ respectively). The median and 98th percentile are computed for the specific location, time of the day and day of the year ±10 d using the available time period. Furthermore, temporal tendencies of p, θ̃ and d are calculated (Δp, Δθ̃ and Δd respectively) and are simply represented by the difference between the current and the previous hour.

As in Part 1 , we concentrate on western and central Europe, more specifically, stations within the area of 40 to 60∘ N and 10∘ W to 20∘ E. After removing stations that measure fewer than three of the five meteorological parameters, around 750 station reports per time step remain on average. For the climatology, we include all time steps from January 2001 to December 2019, i.e. a total of 114 months.

COSMO-REA6 is a reanalysis data set based on the German Consortium for Small-scale Modeling (COSMO) model from the DWD computed by the Hans-Ertel-Centre for Weather Research. The data set covers the European CORDEX (Coordinated Downscaling Experiment) domain with a grid spacing of 0.055∘, i.e. roughly 6 km, and uses ERA-Interim data as boundary conditions. The reanalysis is available from 1995 to mid-2019. For a fair comparison of the two data sets, the chosen time period for the climatology is as close as possible while including 19 extended winter seasons each. This means that, although the observations cover January 2001 to December 2019, the COSMO-REA6 data for October 2000 to March 2019 are used, i.e. with a minor shift of 3 months. In addition to the parameters mentioned above, COSMO-REA6 allows us to include further variables, such as wind gusts at 10 m (vgust), specific humidity at 2 m (q), relative humidity at 2 m (RH) and total cloud cover (cc). The model uses a convection parameterisation by and wind gusts are estimated following and with a turbulent- and a convective-gust component. The cloud cover is based on cloud water and cloud ice; considers grid-scale, sub-grid convective and sub-grid stratiform clouds ; and is verified using ceilometer data . For a detailed description of the data set, we refer to and . The median and 98th percentile used for the normalisation of θ and v respectively, are only computed for the 10-year time period from 2005 to 2015 due to computational cost. The data, originally on a rotated grid, were regridded to a latitude–longitude grid with a spacing of 0.0625∘, i.e. roughly 7 km, for the area of 40 to 65∘ N and 10∘ W to 25∘ E. Note that the area shows an eastward extension and a northerly shift compared with the observational data set to include more northern regions affected by winter storms, where the observational data are sparse.

RAMEFI delivers a probabilistic identification of the five features, WJ, CFC, CJ, CS and NF, with SJ being included in the CJ. The method is based on an RF that was trained on surface observations of 12 storm cases, for which the wind features were subjectively labelled. These case studies are picked to capture a healthy diversity of cyclone developments and features, i.e. they include very intense and more moderate cyclones with differing storm tracks. Given this diversity and the promising evaluation of the method in Part 1 , we assume a reliable detection of the features in long-term data for most cyclones. Note that RAMEFI focuses on strong wind speeds; hence, the RF was trained and tested only for “windy conditions”, which we define as cases with ṽ>0.8. In Part 1 , we used a cross-validation approach for a proper evaluation of the method, i.e. to test the identification for each storm, we trained the underlying RF on the remaining 11 case studies to avoid using data of the storm of interest, resulting in a total of 12 RFs. Here, however, we use the RF trained on all of the 12 case studies . Further, the statistical evaluation of the application on COSMO-REA6 data in Part 1 demonstrates that RAMEFI generates reliable identifications for gridded data despite being trained on surface observations. For details on the method, we refer to .

RAMEFI is spatially independent, such that the output probability is computed individually for each station or grid point. Here, we apply RAMEFI to station observations and COSMO-REA6 data under windy conditions during the extended winter months, regardless of whether a storm occurred or not. However, we later filter the output for cyclone occurrence, as discussed in Sect. . Note that the 12 cases used for training see are also included in the data used for the climatology. The reasoning behind the cross-validation approach in Part 1 was to evaluate whether the RF is able to reliably identify the features in unseen data. Here, we want to generate a climatology of the high-wind features rather than testing the method. Hence, it is unproblematic to apply RAMEFI to the same data that it was trained on. Instead, we obtain an identification that mirrors the subjective identification within in these storms and is still consistent with the entire climatology due to the same model underlying the identification.

In contrast to Part 1 , RAMEFI is also applied to ocean grid points of COSMO-REA6, where it has not been systematically evaluated; thus, results should be treated with some caution there. Considering that the wind speed distribution over sea is broader, the 80 % threshold of the 98th percentile of v results in more windy conditions over the ocean than over land (see Appendix ). Characteristics of other parameters are discussed in Sect. . Nevertheless, looking at various cases over the 19-year period, the ocean and land do not seem to behave fundamentally differently with respect to feature detection and their probability distributions. Exemplary cases can be accessed in the “Video supplement” .

As RAMEFI provides a probabilistic identification, each feature is assigned a probability from 0 to 1. The distribution of the probabilities for each feature is shown in Fig.  for both data sets. While the WJ and CS show a similar distribution in both data sets with peaks around 48 %, the maximum of the CJ slightly differs and is lower at around 43 % for COSMO-REA6 and 40 % for observations. The highest uncertainty in the feature detection can be seen for CFC, which shows overall lower probabilities with a peak around 33 %. The biggest difference in the data sets is found for NF. COSMO-REA6 shows a peak at 50 % but a plateau between 50 % and 73 % in the observations. This will be discussed further in Sect. . This probabilistic information is used in two different ways: firstly, we assign the feature with the highest probability of a given time and grid point, ignoring all other probabilities (referred to as MAXP hereafter). Secondly, we exploit the probabilistic nature of the identification by interpreting the (accumulated) feature probabilities as the expected number of features (referred to as ACCP hereafter). The calibration of the feature probabilities, which was checked in Part 1 of the study, is a critical condition of this approach. The second approach is particularly important for features with less confident detection, which might be underrepresented in the first approach, e.g. the CFC as shown in Fig.  (see also Sect. ).

As an example, consider an identification of 75 % for rain vs. 25 % for no rain. Although we always detect rain via the first approach, rain was actually observed, on average, every fourth case (given that the probabilities are calibrated).

In general, RAMEFI can be used without any filters. However, for a meaningful climatology, we aim to exclude high winds not associated with extratropical cyclones and, hence, the mesoscale wind features are targeted here. Therefore, to filter the gained probabilities and also to compile a storm-relative analysis, objectively determined cyclone tracks are used. The cyclones are identified and tracked from ERA5 p data using an objective tracking algorithm . The algorithm primarily searches for the minimum p in the vicinity of a ∇2p maximum (as a proxy for vorticity) within a radius of 750 km to be assigned as cyclone centres. To filter out weak and thermal lows, or cyclones over high orography, we follow the criteria from . The method settings used for ERA-Interim were slightly adapted to handle the higher-spatial-resolution data of ERA5, while the time resolution was kept at 6 h intervals. Cyclones must travel at least 1000 km and last for at least 1 d to be considered. The resulting cyclone tracks are then interpolated linearly to gain hourly information. The track density, i.e. the number of cyclones passing over a grid point , of all cyclones within the 19 years can be seen in Fig. . As expected, given the selected study area (blue box in Fig. ), cyclone tracks corresponding to the identified features typically travel over the British Isles and the North Sea towards the Baltic Sea.

Here, all grid points showing windy conditions in the vicinity of 15∘ in the zonal direction and -15 to 5∘ in the meridional direction of the cyclone centre and the considered area are used, except those at altitudes above 800 m, consistent with Part 1 . Furthermore, following Part 1, we exclude the Balkans, where topography is complex and where winter storms are rare. For example, in mountainous regions, the Föhn effect might be falsely identified as the WJ. Hence, such areas are removed from the climatology, as indicated by the hatching in Fig. . We further decided to focus on the area of 45 to 65∘ N and 10∘ W to 25∘ E, thereby including the most impacted area over Europe. Note that the changing zonal extent with latitude is neglected in this analysis. However, we do not expect this to have a significant impact on our main conclusions. We remove time steps in which fewer than 5 % of all grid points during the time step are associated with one of the features (excluding NF), i.e. weak cyclones. Time steps with a cyclone moving through and at least 5 % of the area showing windy conditions are referred to as “stormy time steps”. Overall, these filters lead to 1910 cyclones over around 20 000 time steps, which are included in the analysis.

Figure  shows the relative frequency of the wind features for both observations (first row) and COSMO-REA6 (lower three rows) over all time steps and grid points satisfying stormy time step conditions (Sect. ). While the left two columns show all features, including NF, the right two columns neglect NF to focus on the identified mesoscale wind features. Furthermore, the frequencies of the most probable feature (MAXP) are displayed next to the accumulated probabilities (ACCP), i.e. the expectation. For better comparison between the observations and reanalysis, COSMO-REA6 is displayed for land grid points, ocean grid points and all grid points.

As displayed in Fig. a, of the four mentioned features, the CS shows the highest proportion (21.5 %) and is the most probable feature, followed by the WJ (with just under 15 %). The CJ reaches merely 4 %, while the least common feature is CFC with under 1 %. However, NF has a proportion of almost 60 % – almost 3 times as much as the CS. This might be caused by missing values within the data set, which are replaced by the mean values of the variable. This complicates the distinction between the features, as fewer parameters include information about the current conditions, leading to higher probabilities of NF and causing NF to be the most probable feature more often. Indeed, Fig. b shows that the proportion of NF decreases by over 10 percentage points if probabilities for all features and not only the most probable one are taken into account. Although this leads to an increase in all mesoscale features, it is not by the same amount. While the features with overall higher probabilities, namely the WJ and CS as shown in Fig. , increase by around 10 %–30 %, the CJ shows an increase of 73 %. The CFC, which shows the highest uncertainty, increases by 400 %, demonstrating the gain due to using the assigned probabilities. Nevertheless, CFC is by far the least common cause of high winds.

Neglecting NF in Fig. c and d draws the focus to the ratio of the mesoscale features themselves. In the MAXP perspective, the CS and WJ are the cause of high winds in over 50 % and 36 % of cases respectively, while the more damaging CJ and CFC features e.g. only show a proportion of 10.1 % and 1.8 % respectively. However, these are also the features with lower certainty (Fig. ). Hence, from the ACCP perspective, the CJ and CFC come to a total of around 19 %, i.e. an increase of around 60 %. While the WJ also increases slightly, the proportion of the CS decreases by almost 8 percentage points. This suggests that the CJ and CFC mostly lose against the CS with respect to being the most probable feature.

In contrast to the station observations, the proportion of NF is considerably lower in COSMO-REA6 data, as seen in Fig. e and f. This supports the hypothesis that the high proportion of NF in observations is due to missing values, as the data are of course complete for all parameters here. Still, the proportion of NF accounts for around one-quarter to one-third of high winds. This is due to the several reasons discussed below and in Sect.  as well as overall higher uncertainty in uncommon cyclone development, such as double fronts. The reader is referred to Sect. 7 of Part 1 for a detailed discussion . As the overall certainty of NF is lower for COSMO-REA6 and closer to the density of the CS and WJ (solid lines in Fig. ), the proportion in Fig. f actually increases, whereas the CS and WJ proportion decreases. As the proportion of NF also substantially affects the proportions of the other features, we compare observations and gridded data without NF from here on (Fig. g, h). Therefore, apart from the differences in NF described in the previous sentence, the largest difference between the two data sets can be seen in the CJ percentage, with an increase of almost 8 percentage points for land grid points in COSMO-REA6 and over 16 percentage points for ocean grid points for MAXP. This is due to the fact that the CJ commonly occurs in northern continental Europe, over the sea and in Scandinavia, i.e. regions where fewer station observations are available in our data set (see Sect.  and Fig. e). On the other hand, the CS mostly occurs further south, i.e. farther away from the cyclone centre, such that a higher CJ percentage leads to lower percentages for the CS in the gridded data over land. The WJ, which occurs more over land than over sea (cf. Fig. e–h and i–l), shows similar percentages for both the observations and COSMO-REA6 over land. With similar probability distributions for the mesoscale features for both data sets (Fig. ), it is again apparent that CFC shows a higher percentage in ACCP. Comparing sea and land grid points (cf. Fig. e–h and i–l) finally shows that the CJ and CS occur more often and over a wider area over the ocean compared with CFC, which almost exclusively occurs over land, where friction is higher and static stability is lower during daytime. As mentioned before, the WJ is more common over land, around 15 percentage points more prevalent in both MAXP and ACCP.

Finally, Fig. m–p show the proportions for all grid points. Note that the number of land grid points is around 35 % higher than the number of ocean grid points. Overall, almost half of windy conditions are caused by the CS, followed by the WJ and CJ, with around 30 % and 21 % respectively. Again, for CFC, the difference between MAXP and ACCP shows a considerable difference from 1.3 % to almost 5 %. To examine how robust these numbers are, we computed three subsets of nine randomly chosen winter seasons. The proportions vary just slightly with an average of around 2 % (not shown), as is to be expected considering the small fluctuations between winter seasons (see the discussion of Fig.  below).

As the overall frequencies are similar, differences are plausible in both data sets and COSMO-REA6 has the advantage of an homogeneous field without missing parameters, we focus on the gridded data set from here on.

Figure  shows the proportions of each feature for MAXP, analogously to Fig. m and o (COSMO-REA6 all), but only including grid points where ṽ≥1.2 (panels a and b) and ṽ≥1.4 (panels c and d). Note that this is only the case for around 1.5 % and 0.1 % of the previously included data points respectively. Considering only ṽ≥1.2, the proportion of NF decreases by over 75 % (Fig. a), whereas it decreases to almost 0 % if only grid points with ṽ≥1.4 are included (Fig. c). This suggests that higher winds within the vicinity of a cyclone are mostly associated with one of the introduced features. When NF is neglected, the CS proportion is reduced by 45 %–58 % (Fig. b, d); this suggests that, although it affects a large area, it is less common for the CS to be the cause of extreme winds. In contrast, the more damaging features (the CJ and CFC) show an increased proportion. The CJ shows an increase of about 69 % for ṽ≥1.2 and 111 % for ṽ≥1.4. Consistent with , CFC shows an even stronger increase with over 300 % and over 600 % respectively. These results are consistent with the wind characteristics, as will be discussed in Sect. . Meanwhile, the proportion of the WJ decreases by about 10 %, as it usually causes weaker winds compared with the CJ and CFC e.g..

The discussed proportions in Fig.  depend not only on the occurrence of the feature but also on its size. Figure  shows the seasonal and interannual evolution of stormy time steps and the occurrence of each respective feature. This is computed as the sum of stormy time steps during which a certain feature is detected to be the most probable feature in at least 100 grid points over the investigation domain, normalised by the number of stormy time steps. Thus, this measure is independent of the number of grid points in which the feature is identified.

Looking at all stormy time steps of the investigation period (left side of Fig. ), it is evident that NF (grey dot) occurs somewhere in the domain for practically every moment in time. The same holds for the CS (orange dot) with only marginally lower frequencies. Both the CJ (blue) and WJ (red) occur in over 80 % of stormy time steps, with slightly higher values for the CJ. Together with the lower proportion of the CJ compared with the WJ, as seen in all panels of Fig.  except panels i–l (over the ocean), this suggests that the CJ is on average a smaller feature than the WJ. The least frequent feature (with around 43 %) is CFC (green), contributing to the low proportions in all panels in Fig. .

With respect to the mean seasonal cycle, the black line in the middle section of Fig.  shows that November and March are the least stormy months during our investigation period of 2000 to 2019 with a peak in January in between. This is consistent with the cyclone track density plots shown in Fig. . Somewhat surprisingly, October shows the highest number of stormy time steps of all months. On long-term average, the majority of storms occur between December and February; thus, the 98th percentile of wind speeds is highest for that period and lower towards autumn and spring, consistent with and . However, the recent 2 decades shows a larger number of noteworthy storms in October (e.g. Christian in 2013, Xavier in 2017 and Herwart in 2017) compared with November, which leads to a larger number of stormy time steps in the 19 years. This difference might be further enhanced by the slightly lower 98th percentile of v in October compared with November. The higher frequency is consistent with October showing a slightly higher cyclone density than November (Fig. a, b). With respect to the individual features, the rather rare CFC has a marked seasonal cycle with an apparent peak in December and January. The WJ has a smaller relative peak during peak winter months. While the relative frequency of the CJ is only slightly higher than for the WJ, its frequency increases with the winter passing, leading to a maximum of over 90 % in March. Recall that the θ̃ predictor is a normalised parameter, such that a cooling Arctic with a progressing winter and possibly more cold-air outbreaks, i.e. lower values of θ̃, might lead to higher numbers of high-wind events being allocated to the CJ – and to the CS. Both NF and the CS are so frequent that an annual cycle is not evident in this analysis.

Finally, the right section of Fig.  shows the interannual evolution of stormy time steps and wind features. Overall, lower numbers of stormy time steps, such as 2002–2003, 2005–2006, 2009–2010, 2010–2011 and 2012–2013 are consistent with negative values of the North Atlantic Oscillation NAO; e.g., which describes the large-scale circulation over the North Atlantic and originally represents the pressure difference between Iceland and the Azores . For our study area, slightly positive NAO values facilitate stormy conditions, as the typical cyclone paths for such NAO conditions correspond to tracks over the British Isles and the North and the Baltic seas. Accordingly, most peaks are associated with positive NAO phases. Furthermore, quieter winter seasons consistent with the literature, such as 2010–2011 e.g., can be found. With 2009–2010 being a particularly cold winter season , the occurrence of detected WJs is lower compared with other winters, while the CJ shows a peak. The peaks of the WJ and CFC in 2015–2016 are consistent with the winter season being particularly wet and warm, as discussed in . Again, NF and the CS occur too frequently to detect an interannual cycle. Overall, all features have no correlation or a very weak positive correlation with the number of stormy time steps (0 %–10 %). The coefficient of variance is lowest for the CS and stormy time steps (with 41 %) and highest for CFC (with 54 %).

With respect to long-term trends, a slight decline is evident, consistent with the overall decrease in the number of winter storms in a warming climate . However, given that our investigation period covers only 19 years, a Mann–Kendall test significance level of 0.05; did not indicate statistical significance in any of the time series.

An Earth-relative framework enables us to learn which regions are commonly affected by which feature. Figure  shows a geographic distribution of the relative frequency of NF and the four mesoscale wind features (MAXP) as well as of overall windy conditions. An analogous plot for ACCP can be found in the Appendix  showing overall similar properties (Fig. ). Results are displayed on 0.5∘×0.5∘ boxes, thereby aggregating over 16 grid points.

The number of windy conditions, i.e. ṽ exceeding 0.8, in proximity to the cyclone centre (within ±15∘ in the zonal direction and -15 to 5∘ in the meridional direction) is displayed in Fig. a. This criterion leads to the highest numbers over the North Sea, Denmark, northern Germany and the Baltic Sea, i.e. south of the maximum track density shown in Fig. . Furthermore, choosing a threshold of 80 % of the 98th percentile results in more windy conditions over the ocean (see Fig. ); hence, the absolute frequencies of each feature are normalised by the number of exceedances for each grid point. Due to orographic effects and higher noise in these regions, we exclude grid points above 800 m (hatching in Fig. ) and the area east of the Alps, including Hungary, Slovenia and the Balkans. Note that the frequencies of NF, WJ, CFC, CJ and CS add up to 1.

Figure b shows that NF is most common, with a relative frequency of around 50 % in the periphery of the area, i.e. north and further south of the most common cyclone paths and wind footprints. Although areas of high orography are removed, their effects can still be seen upstream when the mostly westerly winds encounter mountain barriers, such as the Scandinavian Mountains, the Western Alps and the Carpathian Mountains, leading to higher frequencies of NF in these regions. Over western and central Europe, high winds are usually closer to the cyclone centre, such that they are mostly associated with one of the mesoscale features.

The WJ occurs mostly over western Europe with relative frequencies of almost 40 % and decreases over Germany (Fig. c) down to 20 % over Poland. This is consistent with an occurrence early in the life cycle when the cyclone is still in the western regions. The higher frequency east of the Scandinavian Mountains should be treated with caution, as it might be caused by orography such as the Föhn effect when a cyclone crosses the mountains. Moreover, a land–sea contrast is visible, which is weaker in ACCP (Fig. ). A possible explanation for this is the different thermal characteristics of the land and ocean, such as daytime heating leading to differences in θ̃. Hence, if just trained over land, RAMEFI might have difficulties distinguishing the WJ from the CS over sea in ambiguous situations. However, looking at exemplary winter storm cases, where features are well developed, a difference in the detection over ocean and land grid points is not evident .

CFC shows low frequencies (under 4 %) for MAXP (Fig. d), while values are twice as high for ACCP (Fig. d) due to the lower certainty of the feature (Fig. ), as discussed in Sect. . However, a distinct land–sea contrast is visible in both, where CFC seems to be almost exclusively detected over land. This might be due to land effects, such as frictional convergence and land surfaces being heated up more strongly than over ocean during the day, leading to the destabilisation of the atmosphere. Moreover, CFC develops slightly later than the WJ when the cold front intensifies (Sect.  and Fig. b). Note the patchy behaviour over land, possibly caused by local small-scale effects due to, among other things, surface roughness and orography. Distinct maxima are found east of or over mountainous regions, such as the Scottish Highlands and the Scandinavian Mountains. Here, again, results given by RAMEFI should be treated with caution. As orography can induce convection, CFC might be detected without the occurrence of a cold front but where high wind speeds are associated with a strong pressure gradient or other features combined with orographic convection.

As expected, the occurrence of the CJ (Fig. e) maximises in the northern half of the domain, much farther north than for the WJ (Fig. c), with a distinct footprint over the northern British Isles, the North Sea and the Baltic Sea. Over the British Isles, where intense cyclones are more frequent than over the North Sea (not shown), the CJ shows a maximum of over 20 %. Overall, the CJ occurs mainly over the sea and coastal areas (although RAMEFI was trained over land).

The CS shows high values in the Bay of Biscay, where other features are rarely detected, and a rather abrupt drop over France in the east (Fig. f). Along with the opposite patterns for the WJ, this suggests a possible false detection in some cases. As explained above, we suspect some systematically different behaviour between the land and ocean to be at least partly responsible for this. A second peak in the CS can be found over eastern Europe, where most other features have already weakened at that late stage in the cyclone life cycle. Overall, the CS occurs further south than the CJ and, thus, farther away from the cyclone centre.

For the system-relative framework, we concentrate on the area within ±15∘ in the zonal direction and -15 and +5∘ in the meridional direction of the cyclone centre. This translates to around ±1073 km in the zonal direction at 50∘ latitude and 1670 and 557 km in the southern and northern directions respectively. Figure  shows a composite over the 19 extended winter seasons from 2000 to 2019 relative to the cyclone centre (Fig. a) and life cycle (Fig. b).

With respect to the mean spatial distribution, the WJ mostly occurs within the south-eastern quadrant of a cyclone, consistent with conceptual models see Fig. 1 in Part 1;. As shown in Fig. a, the WJ usually has a distance of 250 to 1500 km from the centre. CFC occurs around 3–5∘ farther to the west, i.e. upstream with respect to a westerly flow and also slightly shifted to the north, closer to the cyclone centre. As CFC is a relatively small elongated and narrow feature (as is the front itself), the location is harder to pinpoint over so many cases, and the location varies the most from case to case compared with the other features. Thus, although CFC overlaps with other features statistically, this should usually not be the case for individual cyclones. The CJ is situated south to south-west of the cyclone centre with some statistical overlap with CFC. It occurs closer to the centre than the CS, which dominates the south-west quadrant and can have a distance of up to 1500 km from the cyclone centre. The 25 % shaded area for NF is split into two patches. The majority of NF is detected south to south-east of the cyclone centre, mostly coinciding with the WJ but extending its reach to the north and south-west. The northern part of this patch is located in the area of the warm front and is possibly connected with the CCBa, as discussed in and . A second, smaller patch can be found to the north-west of the cyclone centre. In this region, the CJ usually occurs before it is wrapped around the cyclone centre. As the CJ follows the bending of the front, the wind direction differs from the CJ later on, when it wraps around the centre, such that RAMEFI does not identify this part of the CJ. However, high wind speeds in this area are usually only caused by very strong CJs, as the relative movement of the air in this area is against the cyclone motion, weakening the Earth-relative wind speed, particularly for fast-moving weaker cyclones . As shown in Appendix , the relative frequency of the features shows that NF mostly occurs in the north-eastern quadrant, where other features are rare and where windy conditions are less common (Fig. ).

Figure b shows the relative frequency of the four mesoscale wind features throughout the lifetime of the parent cyclone from 2 d before until 2 d after the time of maximum depth, i.e. the deepest pressure minimum during a cyclone's life cycle (marked as 0 in Fig. b). The WJ is the first feature to develop, with a maximum at -6 h. A total of 50 % of the detected WJ grid points occur between -18 h and the time of maximum depth. With a small offset of around 2–3 h, the CFC follows, consistent with its more western location in Fig. a. The peak is slightly lower at around -4 h, and 50 % of CFC points are detected between -16 and 6 h. Contrary to , the CJ develops several hours before the time of maximum depth. However, the peaks of CJ and CS are around the time of maximum depth, even though the peak of CS is lower. The WJ and CFC have already began to decrease at that time, as the warm air begins to occlude, consistent with . While the CJ occurrence decreases faster, the CS remains for longer. Overall, 50 % of the CJ and CS events occur within -8 and 10 h and within -8 and 14 h respectively.

When combining both composites, we can see how the features occur in different locations during the cyclone's life cycle. This is displayed for four exemplary time slots in Fig. . An animation showing all time slots from 24 h before to 24 h after the time of maximum depth can be accessed in the Video supplement . A total of 24 h before the time of maximum depth, only the WJ and CFC appear in the composite: the round maximum of the WJ located south to south-east of the cyclone centre and the CFC with a more north- to south-elongated maximum closer to the cyclone centre (Fig. a). The CJ and CS develop in the following hours to the south-west of the cyclone centre, while the size of the WJ increases, as shown 12 h later (Fig. b). The area of CFC also increases, but the increase in width is probably rather due to the variation in location than an increase in size. Around the time of maximum depth, as shown in Fig. c, the WJ area decreases and shifts farther away from the centre in a south-eastern direction, now with a stronger west–east orientation. The CJ has increased in size and now stretches across both southern quadrants, whereas the CS fills most of the south-eastern quadrant. Furthermore, NF covers most of the area overlapping with all other features. As mentioned above, the area north to north-east of the cyclone centre, which does not overlap with any of the mesoscale features, corresponds to the CCBa, as described in and . A total of 12 h after time of maximum depth (Fig. d), the WJ and CFC have mostly vanished, while the CJ and CS are much diminished in size.

Overall, these results are mostly consistent with idealised schematics and conceptual models in the literature e.g.their Fig. 1. However, the very large set of differing cyclone development information in our comprehensive data set is able to show a larger variety. As our study domain is too small to cover the whole life cycle of the investigated cyclones, especially the early stages of a feature might be missed, such that an analysis of feature duration and comparison with the literature e.g. is not meaningful. Over the investigation domain, the WJ, CJ and CS have a broadly similar duration, with an average of around 20 h (not shown). In contrast, the smaller and rarer CFC appears for only half as long, with an average of around 11 h (not shown).