WCD-2021-18 Identification , characteristics , and dynamics of Arctic extreme seasons

The Arctic atmosphere is strongly affected by anthropogenic warming leading to long-term trends in, e.g., surface temperature and sea ice extent. In addition, it exhibits a pronounced seasonal cycle and strong variability on time scales from days to seasons. Recent research elucidated processes causing short-term extreme conditions in the Arctic that are typically related to the occurrence of specific weather systems. This study investigates unusual atmospheric conditions in the Arctic on 5 the seasonal time scale, characterized by surface temperature, surface precipitation, and the atmospheric components of the surface energy balance. Based on a principle component analysis in the phase space spanned by the seasonal-mean values of the considered parameters, individual seasons are then objectively identified that deviate strongly from a running-mean climatology, and that we define as extreme seasons. Given the strongly varying surface conditions in the Arctic, this analysis is done separately in Arctic sub-regions that are climatologically characterized by either sea ice, open ocean, or mixed conditions. 10 Using ERA5 reanalyses for the years 1979-2018, our approach identifies 2-3 extreme seasons for :::: each :: of : winter, spring, summer, and autumn, respectively, with strongly differing characteristics and affecting different Arctic sub-regions. While some show strongly anomalous seasonal-mean values mainly in one parameter, others are characterized by a combination of very unusual seasonal conditions in terms of temperature, precipitation, and the surface energy balance components. Two ex15 treme winters affecting the Kara-Barents Seas are then :::: Kara :::: and :::::: Barents ::::: Seas ::: are selected for a detailed investigation of (i) their substructure, (ii) the role of synoptic-scale weather systemsthat occur during the season, and (iii) , :::: and potential preconditioning by anomalous sea ice extent and/or sea surface temperature at the beginning of the season. Winter 2011/12 shows the highest surface temperature anomaly in parts of the Kara-Barents Seas (about + 5 K), which was due to :::::: started :::: with :::::: normal ::: sea :: ice :::::::: coverage ::: and :::: was ::::::::::: characterized ::: by constantly above-average temperatures during the season related to a strongly :::::::: sequence 20 :: of ::::::::::::: quasi-stationary ::::::: cyclones ::: in ::: the :::::: Nordic :::: Seas, :::::::: favoring ::: the ::::::: frequent :::::::: advection :: of ::::: warm ::: air :: to ::: the ::::::: Barents :::: Sea. ::: An enhanced frequency of blocking anticyclones in the Kara-Barents Seas and a strongly ::: and :: a reduced frequency of cold air outbreaks . Sea ice coverage was normal at the beginning of the season and then developed a negative anomaly due to the unusually high temperatures :: in ::: the :::: Kara :::: and ::::::: Barents :::: Seas :::::: further :::::: helped :: to :::::::: maintain ::: the ::::: warm :::::::: anomaly. In contrast, winter 2016/17 started with a strongly negative anomaly in :::::: reduced : sea ice coverage and a strongly positive anomaly in sea surface temperature 25 in the Kara-Barents Seas, which remained during most of the season. The combination of this preconditioningwith specific synoptic conditions, i.e., a particularly high frequency :::::::: enhanced ::: sea :::::: surface ::::::::::: temperatures :: in ::: the ::::: Kara ::: and ::::::: Barents ::::: Seas. :::: This

This research nicely demonstrates how PCs can be used to identify seasonal anomalies and extremes in certain regions of the Arctic. It furthermore demonstrates how to use that information to provide an expectation of how an extreme season was characterized with regard to one of the six variables and how consistent those conditions were. It is certainly a nice way to be able to identify extreme seasons that might be worth analyzing in further detail at shorter time and space scales if desired. Overall, I think these results can make a contribution and be published once some remaining issues are addressed. In particular: 1) Picking the first two principal components is subjective and does not necessarily isolate most of the variance. It first needs to be established that the first two principal components are the only significant ones. I do not doubt that this is the case given that on line 282, it is stated that they usually explain 80-90% of the variance. However, it should be shown that they are indeed statistically distinguishable from the others.  provide a well-established method of statistically distinguishing the first few eigenvalues from the others.
Thank you for this comment and for pointing us to the method introduced by North et al. We applied this method and the results reveal that the first two PCs in DJF and JJA are, with the exception of sub-region ARM in JJA, always statistically distinguishable from the others. We added this information to the revised manuscript (L305ff.). Here we provide further details about our results from applying the  method, which we also show in the supplement. Figures R1 and R2 show the standard errors for each eigenvalue in our PCA as introduced by  for each sub-region in DJF (left-hand side) and JJA (right-hand side), respectively. The estimate for the standard error is given by , δλ α ≈ λ α (2/ ) 1/2 where denotes the respective eigenvalue and N the sample size, which in our case λ α corresponds to the 39 realizations of the four seasons of the study period. Along the y-axis, the eigenvalue for each Principal Component (PC) is given and the error bars represent the estimated standard error. For both seasons, the first two eigenvalues, and are either clearly The only exception is the sub-region ARM in JJA for which the error bars of and have a λ 3 . λ 2 λ 3 significant overlap. Thus, we can show that the first two PCs, which we use for the definition of our extreme seasons and which explain between 80%-90% of the variance in the respective sub-regions, are almost always statistically distinguishable from the remaining eigenvalues. We conclude that PC1 and PC2 isolate most of the variance and the corresponding eigenvalues are statistically distinct. The number of the Principal Component is given along the x-axis and the eigenvalue of each Principal Component along the y-axis. Error bars denote the estimated standard error following .
2) Section 3 and generally throughout: The values of all correlations and their p-values that are described should be listed in a table.
We agree that it would be helpful to add a list containing all correlations and respective p-values for the described relations between the different parameters. We thus added such a table to the supplementary material and refer to it in the paper.
3) Figures 5 and 6 are a very nice way to illustrate the seasonal anomalies and the variability that may have also been occurring within those seasons. Having never seen these diagrams before, it at first takes a little bit of time to understand. It would be very helpful if there were a schematic showing the "phase space" of the interpretation that illustrates what is said in words on lines 251-258 (i.e., regions on the graph where there would be anomalies that tend to be continuous, where there would be warm episodes alternating with weak cold episodes, where there would be several intense warm and cold episodes that nearly cancel, where they would be near the climatology, etc.).
Thank you very much for pointing this out. To better understand and interpret the figures, we added lines of a constant ratio of the seasonal-mean anomaly and the seasonal-mean absolute anomaly ( ) to the diagrams, such as you can see in the schematic figure below. We further 'To better understand the seasonal substructure of Arctic winters and summers, we compare the seasonal-mean anomalies ( ) with the seasonal-mean absolute χ * anomalies ( ) for T 2m , P and E S in selected sub-regions and negative daily anomalies cancel each other, leading to a weak overall seasonal anomaly.
The value of is further indicative of the magnitude of the daily anomalies throughout a χ * | | season. A season located at the top of the plot shows stronger daily anomalies than a season with the same ratio but a smaller .
For example, a season can be anomalously warm because the daily-mean T 2m values are larger than the climatology on almost all days of the season , resulting in . With a decreasing 1) Line 135: Why are only marine cold air outbreaks (CAOs) considered? There are also significant CAOs over land, described in Biernat et al. (2021).
We are only considering ocean and ice grid points and thus only marine cold air outbreaks, which are identified on grid points with less than 50% sea ice.
2) Line 186: Choosing a dM threshold of 3 seems quite subjective. How is this threshold picked? If each principal component has a significant anomaly of two standard deviations, this could provide an expectation for what would be significant when considering the PCs in combination.
The thresholds for anomalous resp. extreme seasons are indeed a rather subjective choice. However we find that with these thresholds we obtain on average 0-1 extreme seasons per sub-region (which equals 0-2.5% of all seasons) and 4-5 anomalous seasons per sub-region (equalling 15-17% of all seasons). Assuming a normal distribution, these values correspond to the range of 2-3 for our extreme seasons and 1-2 for our anomalous seasons. Further, with σ σ this number of extreme seasons, the return period of such a season corresponds to approximately 40 years. Several studies, e.g.,  used this as an adequate measure for defining extreme seasons.
As a side note, we would like to mention that preliminary analyses of 1000 years of (present-day) CESM large-ensemble data show that our chosen threshold of d M =3 results on average in a return period of around 70-90 years. We are, thus, confident that classifying the seasons with d M >3 as "extreme" is well justified.
3) Line 219: Be more specific about "almost always." What percentage of the time is it true? Same thing for line 225... what percentage of the cases translates to `usually'?
Thank you for pointing this out. We adjusted the manuscript in the indicated section (L225ff.) to clarify the mentioned relationships between the different variables. 4) Line 262: How close to the |P*| = P* line does a season need to be in order to be called "continuous?" For example, the 2016/2017 winter season was pretty close, but not exactly on it. On the other hand, there are very few cases of |T2m*| = T2m* being exactly equal in the summer while it is described as "continuous" on line 260.
There are indeed only very few cases where the seasonal-mean and seasonal-mean absolute anomalies of a season are equal. Thus we changed our definition of a "continuous anomalous season" from to , including seasons with a ratio of (see previous χ * | | = χ * χ * | | ≈ χ * 0. 8 ≤ χ * χ * | | ≤ 1 comment). We added these changes to the revised manuscript (L265ff.). 5) Line 307: Would also be useful to point out that there is very little 2-m temperature variability over the Arctic sea ice in the summer. This could imply that temperature variability may not play a major role in sea ice loss, which has very large interannual variability in the summer. Figure 8 does indeed suggest that T 2m has only little variability in regions with SIC clim >0.9 in summer compared to other sub-regions. However, T 2m is capped above sea ice as the air is cold and the excess energy goes into the melting of the ice if T 2m is above the freezing point, which essentially limits (near-surface) temperature variability. We further assume that the sea ice loss in summer is equally strong in the other sub-regions (especially the mixed sub-regions with very variable SIC), which show a larger variability in SIC. As there exists only one sub-region with SIC clim >0.9 in summer, we think that additional analyses would be needed to make such a statement.
6) The justification of how an extreme season is chosen on Lines 310-314 should be moved up to Section 2.3.
We already explain this in Sect. 2.3. The text in lines 334ff in the revised manuscript is meant as a reminder. We now clarify this by writing "As explained in Sect. 2.3, …". 7) Line 315: Which season does Figure 2 show? This could also be referenced here along with Table 2.
Figure 2 in the manuscript is only a schematic biplot which does not refer to a specific season nor region. We slightly changed the figure caption to clarify that this is only an idealised plot. 8) Figures 9, 10, 14: Would be helpful to label the x-axis with the month/date instead of the day of the season, esp. to be consistent with the text.
We mostly use "on day 12,15, 20…" throughout the text and only rarely real dates. Thus, we adapted the manuscript such that we don't use specific dates anymore, as we think that this ensures better readability. 9) Line 367: How are blocking, cyclone, and CAO frequencies computed exactly? Need references and a short description.
A common feature of our weather system identification schemes is that they produce a two-dimensional binary field, often referred to as the "mask" of the weather systems, where grid points that belong to a system have a value of 1 and the others have a value of 0. Simple time averaging of these binary fields then automatically delivers the weather system frequency field. For example, if a cyclone mask covers a grid point at 25% of all times, then averaging 25% times a value of 1 and 75% a value of 0 leads to a frequency of 0.25 (we added this information in lines 397ff.). For the specifics of the identification scheme, we added a few sentences for each weather system and now give the relevant references to the papers that introduced these schemes in lines 132ff. 10) Line 389: "Several episodic precipitation events..." But wouldn't Fig. 5h suggest consistent precipitation events?
Thank you for pointing this out. We deleted "episodic" to clarify the constant occurrence of precipitation events throughout the season. 11) Line 431: Remove "it is obvious that" We changed "it is obvious that" to "it can be seen that". 12) Line 440: Please also label JJA 2016 in Figs. 6 and 8.
To make it clear that JJA 2016 is somehow connected to our case study DJF 2016/17, we additionally labeled it in Figs. 6 and 8 in the manuscript.
13) Lines 441-445: It is misleading to say that there were positive temperature anomalies over large parts of the Arctic in JJA 2016. This and the blocking was more centered over the Kara-Barents Sea region, while much of the central Arctic was not exceptionally warm and had frequent cyclones.
We are not sure if we understood your remark correctly, as we do not state that the positive surface temperature anomalies in JJA 2016 occurred over large parts of the Arctic, but only in the Kara and Barents Seas. We then state that in autumn 2016 (mainly during October and November), positive temperature anomalies occurred across the whole Arctic region as already shown by Tyrlis et al. (2019). We now clarify this further in the text by replacing "during autumn 2016" by "during October and November 2016" (L482).
14) Section 5.3: If the blocking frequency was greatest over Scandinavia, why were the warmest temperature anomalies over the Kara-Barents (KB) region and not co-located with the blocking? Seems like there should have been northerly flow over much of the KB region from air flowing over sea ice. Is it surprising that the air mass was not modified by the time it reached KB?
Thank you for pointing this out. First of all we want to emphasize that DJF 2016/17 was not a particularly warm season, but experienced several episodic warm events. Blocking over Scandinavia influenced the surface temperatures in the Kara and Barents Seas, especially during the warm episode in February 2017. Trajectories show that a majority of the air causing this warm episode originated over Scandinavia and was undergoing subsidence (we will add a short evaluation of some air mass trajectories to the supplement; see answer to comment (15) of reviewer 2 and Fig. R6). However, the pattern of blocking and cyclone anomaly patterns as shown in Fig. 12 in the manuscript does also support northerly flow into the region as you correctly assume, causing for example the period with a strong CAO in mid-February 2017, when cold air is transported from the High Arctic towards the South, facilitated by a block over Scandinavia and a cyclone in the eastern part of the Kara and Barents Seas. Please have a look at the supplementary animation S2 where we show the synoptic evolution for each day throughout the season. We also tried to further shape section 5.3 to better highlight in which way the synoptic patterns influenced the surface temperatures in our case study sub-regions.
15) Lines 120, 541: Is this approach really novel given that ) first introduced it in a similar application?
Using a PCA for finding dominant variability modes has been done in several studies such as, e.g., Graf et al. (2017). However it has never been used to define anomalous or extreme seasons based on the combination of several parameters. Thus, in terms of defining extreme seasons, this approach is novel. However, we deleted the word "novel" in L584 to clarify that the use of a multivariate approach per se is not novel.

Technical corrections:
1) Table 1: 2 m temperature --> 2-m temperature We followed the WCD submission guidelines (see "House standards" for hyphen usage: "It is our house standard not to hyphenate modifiers containing abbreviated units (e.g. "3-m stick" should be "3 m stick")).
2) Table S1: Caption states standardized values are in brackets, but they are instead in parentheses.
Thank you for spotting this. We replaced "brackets" with "parentheses" in the caption of Tables 2  and S3-S6. 3) Section 2 should be "Data and methods" given that there is more than one method used to complete the analysis.
Changed as suggested. Figure 1 caption: State what the green and red boxes denote.

4)
We added the following sentence to the caption of Figure 1: "Green and red boxes denote the areas of the Kara and Barents Seas and Nordic Seas regions, respectively." 5) Line 135: There does not need to be a space between the number and the "%" symbol Again, we followed the WCD submission guidelines (see Figure content guidelines: "Spaces must be included between number and unit (e.g. 1 %, 1 m).").
6) Lines 140-141: What is the sign convention for the surface energy balance?
Thanks for hinting at this. We added the following sentence in L144ff.: "Positive signs denote energy fluxes into the surface, whereas negative signs are indicative for energy fluxes into the atmosphere." 7) Line 183: There should be a period at the end of the equation.
A period has been added at the end of the equation.

Reviewer 2
The paper presents an analysis of variability in three Arctic regions using 6 metrics. An input of those metrics into the dominant modes of variability and links between those metric are discussed. Overall, I am impressed by the amount and quality of work done in this study.
Here is what I like about the paper: • Fig 5 and 6, which show that while strong anomalies may be observed in one or two metrics, other metrics may remain close to their climatological values; • assessment of the input of the six metrics into the main modes of variability and relationships between them; • case studies (particularly fig. 10,11, 14) and the discussion around them. An attempt to establish a connection between the weather and seasonal anomalies is valuable; • a wide range of metrics used in the study -not only T2m/SIC/P, but also energy fluxes, cyclone frequency, CAO and a blocking index.
However, there is a couple of major concerns that need to be addressed before the paper can be accepted for publication: 1. I am not convinced that the approach, introduced in the paper, is a good way to select extreme seasons. Despite the use of a multivariate approach, it often comes to just one metric showing a strong seasonal anomaly, which was enough to identify the season as extreme or anomalous. Thus, without applying this approach, one may simply go through all 6 metrics and select the most extreme season(s) in each of them. I don't think I saw a proof that the seasons selected with the PCA analysis were more anomalous than those that showed a strong anomaly but were not picked up by the PCA approach. The latter may be even more anomalous than those, that were selected using the PCA.
On the other hand, there are seasons that were identified as anomalous though none of the variables showed a strong anomaly. Could it be proved that they are 'true' anomalous seasons and not artefacts of the method?
I am not asking for a change of the approach here, but I think more discussion around potential (dis)advantages of the proposed method is needed. In my opinion, this method identifies the dominant modes of variability and allows for assessment of the contribution of each of 6 metrics into those modes and a link between them. Section 5 explores a few seasons when one of the first two modes of variability was among the strongest.
We appreciate and fully understand this remark. It is a priori not clear how extreme seasons should be defined. An obvious choice, which we also use in other studies, is to simply choose the warmest or wettest seasons. This would prioritize one parameter (e.g. temperature or precipitation) and a justification would be given why this parameter is particularly relevant. Here we tried something else, something more "objective" in the sense that we did not want to pre-specify the most relevant parameter. Instead, we allow for the possibility that besides individual parameters, also their combination can be unusual. Thus, we were led by the hypothesis that our multivariate approach will lead to different types of extreme seasons (different in terms of their individual anomalies of T 2m , P and E S ), which, however, share a similar "anomalousness" as expressed by the parameter d M . We don't think that this method produces artefacts; in order to reach a value of d M ≥2 (or even d M ≥3), at least one of the considered variables or a combination thereof must be clearly exceptional compared to the other seasons in the ERA5 time period. In the revised version we will make sure that this line of thought becomes obvious to the reader. At the same time, we cannot (and don't want to) prove that this approach is "better" than a more conventional one. If all that matters in a specific study is for instance the seasonal snow accumulation, then there is no need to work with our approach.
We adapted the discussion (L584ff.) to clarify the novelty and characteristics of our approach in the context of other, more conventional, methods.
2. My other concern is the length of the manuscript. Considering the amount of work, it is hard to make it shorter, but I think the paper will benefit from it. Some plots (especially, Fig. 3) are too busy and are difficult to interpret. Section 3 and 4, while interesting, are hard to read, particularly when plots discussed in the text are a couple of pages away (which is inevitable). Please select the most robust and/or important relationships and focus on them. I understand that each plot provides a lot of information, but, unfortunately, human beings can only keep a few facts in mind at a time.
We did our best to further streamline the text and make it as readable as possible. With regard to the length of the paper, we think that it is still fairly OK.
Other comments: 1) Abstract: 1.The abstract is a bit long, even if there is no word limit, a page-long abstract is not ideal.
Thank you, we shortened the abstract by about 20%.
2) I think it is worth mentioning that 2016/17 winter was mostly anomalous in terms of precipitation and maybe in some other variables, otherwise, until you read the paper, it remains unclear why it was anomalous.
Thank you for this remark. In the revised version of the abstract we write "In contrast, winter 2016/17 started with a strongly reduced sea ice coverage and enhanced sea surface temperatures in the Kara and Barents Seas. This preconditioning, together with increased frequencies of cold air outbreaks and cyclones, was responsible for the large upward surface heat flux anomalies and strongly increased precipitation during this extreme season." This makes it clear that DJF 2016/17 was mainly anomalous in terms of precipitation and surface heat fluxes.
3) Sect. 2.3: For the PCA analysis, was each metric first averaged over the corresponding region? Meaning that the special structure of those anomalies was not accounted for.
Yes, we average over the region and therefore lose information about the spatial structure. 4) Fig. 3: As I already mentioned above, it is a very busy plot, which is hard to read. The only thing that is obvious to me is that in JJA the red/blue markers can be linked to positive/negative temperature anomalies. For DJF, what is obvious is a link between T2m and P anomalies and that the low right corner has predominantly negative Es anomalies. However, regional differences, discussed in the text, are very hard to see. If you decide to keep this plot, maybe splitting into different geographical locations or the sea ice concentrations helps.
Thank you very much for the suggestion. We adapted the figure (Fig. R4 below shows the revised Fig. 3 of the paper), and now show the correlations for each SIC clim range in separate panels. Figure   In order to limit the length of the paper (see also your remark above), we selected some sub-regions for Figs. 5 and 6 (those that are most relevant for the case studies). The other sub-regions are now shown in the Supplement. 8) l.285-287: The statement on correlation between T2m and P comes from the fact that the corresponding blue lines are close to each other?
Yes. The more parallel two precursor vectors are, the stronger is the correlation of the precursors, provided that PC1 and PC2 explain a large part of the variance . 9) Regarding the comment on the weather systems creating extremes in the high Arctic, I would like to agree, though none of the AR seasons across all regions in fig. 5 look particularly extreme. How about other regions that have stronger extreme seasons often just in one parameter -can they be explained by anomalous weather patterns?
Yes, we agree, variability is distinctively smaller in the High Arctic in winter. But we think that one strength of our method is that it is still able to objectively quantify the "anomaly magnitude" of one season compared to another in a specific sub-region. A method using absolute thresholds would find most extreme seasons most likely at the poleward end of the storm tracks.
With respect to your 2nd question, we unfortunately don't understand what you mean by "other regions". For all seasons, which we investigated in detail, we found an important role of anomalous weather patterns.
10) l.303-307: Why the described connection between P and Rs over the sea, as well as between T2m and RL in KBM does not hold in the Kara-Barents Sea?
In NOS and ARS we can see an anti-correlation between P and R S (corr(P,R S )=-0.91 in NOS and corr(P,R S )=-0.96 in ARS. Our argument for this anti-correlation is the presence of clouds during rainfall. As you correctly point out, there is no such anti-correlation in sub-region KBS (corr(P,R S )=0.05). This would indicate that there is a large variation in cloud cover (and thus possible reduction in Rs) also during periods without precipitation. We did, however, not investigate this relationship in further detail.
We also want to point out that the mentioned correlations are only an approximation (see previous comment), which is more accurate the more of the total variance is explained by the first two Principal Components. As in the mentioned sub-regions this explained variance ranges between 85% and 88%, we assume that the correlations are good enough to use them for our interpretation.
11) Section 4: The relationship between 6 metrics during cold and warm seasons, gained from the PCA analysis, is interesting. Could correlations found in this section be confirmed by using the raw data?
We are not sure if we understand this question correctly. Yes, we can confirm some of the correlations found with the PCA. For example it is shown in Fig. 3 that in DJF T 2m and P are mostly positively correlated in ice and mixed sub-regions, whereas there is no such correlation in regions over the open ocean. This correlation can as well be seen in the PCA biplots for DJF in Fig. 7. However, as we use six different variables for the PCA analysis, and there seems to be no conventional method to illustrate the correlation of six variables, we only show T 2m , P and E S in Fig. 3. It is further important to mention that we use detrended seasonal-mean anomalies for the PCA and thus remove the seasonality and a potential trend compared to the raw data. Therefore, it is not straightforward to compare the correlation of certain parameters for both data sets.
12) l. 322: "By design, extreme seasons have very large anomalies for at least one parameter… However, some anomalous seasons don't show very strong anomalies in one particular parameter, which implies that for these seasons it is the combination of several parameters that makes them anomalous" I am not sure that the first sentence is true. Moderate anomalies in a few variables may also give an anomalous season and this is what happens in some cases.
Thank you for this remark. We rephrased the mentioned section (L349ff.) in order to clarify that indeed our extreme seasons have at least one large anomaly in one parameter (see previous comment about the approach as well as answer to comment 1 by the Editor) to reach a d M value which is larger than 3. However, it is correct that this is not necessarily the case for anomalous seasons, where it is often the combination of several moderate anomalies resulting in d M >2 (but smaller than 3). 13) l.367: I could not find a description of how cyclones, CAO and blocking events were defined.
For the specifics of the identification scheme, we added a few sentences for each weather system and now give the relevant references to the papers that introduced these schemes (L132ff., see answer to specific comment (9) by reviewer 1).
14) l.372: Even during CAOs the temperature remained above the climatological mean, hence, I doubt that 38%-deficit in CAO can be responsible for the season being anomalous. During the first month (days1-27), there were no significant blocking events and CAOs, but T2m was well above average. To me it looks like there was a strong preconditioning. Furthermore, in the next case, shown in Fig. 10, there is a high number of CAOs but they have relatively small effect on T2m, especially during the first half of the season, brings the temperature down by only, perhaps, 2-3 deg.
Thank you for this remark, it is certainly important to discuss this more thoroughly. As we state in section 5.3, where we discuss the synoptics throughout the winter 2011/12, one important feature is the pathway of the cyclones entering the Arctic from the North Atlantic, as they tend to slow down and get stationary in the region of the Nordic Seas, and their position relative to the Kara and Barents Seas. As a result, during several days of this winter, the warm sector of a cyclone is located in the Kara and Barents Seas whereas its cold sector is positioned in the Nordic Seas. This does not only explain partially the relative lack of CAOs, but also the overall increase in the surface temperature anomaly. If the cyclones were located further east, both the warm and the cold sectors would have been located in the region, likely resulting in no notable T 2m anomaly. Comparing the timeseries in Fig. 9 with the supplementary animation S1 shows that this synoptic situation especially occurs in December and in the second half of February, when the T 2m anomaly is very strong. For further studies it could thus be very useful to have a metric for the coverage of a region by a cyclones' warm sector as opposed to its cold sector (and thus the position of a cyclone with respect to that region). This would simplify the interpretation of a cyclones' influence on surface parameter anomalies in a distinct region. In the revised manuscript we now emphasize more that the impact of cyclones depends critically on their track relative to the region (L500ff.).
With regard to your comment on preconditioning in this season, we can say that this is most probably only a minor reason for the anomalous surface temperatures. Indeed, SON 2011 shows already slightly positive values and a slightly reduced SIC, but not to an extent that 2 * could explain the strong seasonal-mean T 2m anomaly during DJF 2011/12. The sea surface temperature reaches values of about +1-1.5 K above normal in September 2011, however returns to climatological values in October and shows no significant anomalies throughout November.
15) l.465: A seasonal blocking anomaly over Scandinavia is probably not enough to support the statement that 'Subsidence-induced warming [over Scandinavia] and long-range transport of warm air masses contributed to several warm episodes.' This is indeed correct. To confirm this statement, we added a short evaluation of some air parcel trajectories to the supplement, which show the importance of subsidence-induced warming and long-range transport during episodic warm events in DJF 2016/17. Figure R6 shows air parcel trajectories for two warm episodes in DJF 2016/17 from 16-19 January 2017 (Fig. R6a) and from 11-14 February 2017 (Fig. R6b). In January, the influence of long-range transport of air parcels at lower levels, mainly from eastern Europe, can be observed. In February, subsiding air masses, favored by the presence of a blocking system over Scandinavia, additionally contribute to the warm event. 16) l.498: why a persistent high does not cause subsidence warming? and why there are no blocking events during Jan 2013 at the time of a persistent high? I can also see a number of cyclones in Feb, despite the text says that Feb was also calm. I agree that probably the main reason for decreasing t2m and low P is that the High Arctic remained isolated from the lower latitudes, however, none of the metrics in this study reflect an exchange between latitudes. I am not suggesting adding such metric at this stage, but it might be something to add in the future.
Thank you for these remarks. It would certainly be useful to have a measure which is indicative for latitudinal air mass exchange to better understand the processes leading to extreme seasons in the High Arctic.
Regarding your questions about the non-co-occurrence of the persistent high-pressure system as well as the lack of subsiding air, we analysed the geopotential height as well as the potential vorticity (PV) anomaly at upper levels throughout this winter. Figure R7 shows the geopotential height at 300 hPa (Z300) during the episode of the strong high-pressure system between 15 January and 25 January 2013 in the region of the Chukchi Sea and the High Arctic. Z300 does not show significantly enhanced values above the surface high, indicating that there is no strong upper-level forcing in the form of a persistent ridge which could have caused the formation of a block and the strong subsidence of air. The analysis of the vertically averaged potential vorticity anomaly (VAPVA) between 500 and 150 hPa does further support these results, as it reaches only small negative or even positive values in the same region (for the identification of a block following , an area with VAPVA < -1.3pvu which persists for at least 5 days would be needed). Thus we assume that the strong high-pressure system at the surface is caused by very cold air below an inversion layer, decoupled from the synoptics in the upper troposphere. We can show that there exists a strong inversion layer very close to the surface in the center of the high pressure system by using a skewT-logP diagram (see Fig. R8), which supports our assumption that the air in this area experiences radiative cooling opposed to subsidence-induced adiabatic warming which one might expect in the presence of an upper-level block.  17) l.529-534: the paragraph first describes obvious seasonal differences (higher variability in winter due to stronger gradients) and then concludes 'hence, it is reasonable to subdivide the Arctic into several regions considering these spatial differences to study anomalous Arctic winter seasons.' But during summer the regions were also subdivided. I am not sure if this paragraph is needed at all.
Thank you very much for this remark. The mentioned paragraph is indeed a bit misleading and possibly not needed at all, which is why we deleted it. However, we still want to mention the difference in spatial variability between winter and summer and therefore added a sentence in this regard to the previous paragraph (L568).

18) l. 541: see my major comment on the PCA approach
See our response on p.12 of this reply document.
Minor comments: 1) l.61 'and of the feedback': remove 'of' Changed "strongly affect the type of linkages between parameters and of the feedback processes" to "strongly affect the type of linkages between parameters as well as feedback processes".
2) Table 1: Es should be added Thank you for pointing this out, we added the variable "Es" to Table 1.
3) Table 2 is first mentioned in section 2.3 but is only shown in section 4. Replace 'brackets' with 'parentheses' Thank you, we changed "brackets" to "parentheses" in the Table caption.
Indeed we refer to Table 2 already in the method part to justify the detrending of our data set. However we prefer to show Table 2 only in the results part and not yet in the methods part as it basically shows the results of our analysis, based on the PCA biplots in Figs. 7 and 8. 4) l.160: it is not the entire ERA5 period, but the entire period covered by this paper Thank you for pointing this out. We changed the regarding sentence to "A distinction is made between areas where, on all days of the considered season in the time period covered by this study, mainly sea ice is present …".

5) Please use either the Kara-Barents sea or the Kara and Barents seas
We now only use "Kara and Barents Seas".

7) 432: on this date
Thank you, changed as requested. Regarding days/dates: see our answer to specific comment (8) of Reviewer 1.
We added SLP to Figs. 10 and 11. Further we removed the CAO heatmap description from the caption. It does not make sense to show the marine air outbreak frequency for sub-region ARI, as this region is mainly ice-covered and as mentioned in the method section (L139ff.), we define CAOs only for grid points with a sea ice concentration of less than 50%.

Reviewer 3
The authors have investigated seasonal extremes in the Arctic using PCA of six climate variables and analysis of some key dynamical elements -cyclones, blockings, and marine cold air outbreaks -to further investigate particular extreme seasons. This is an interesting and valuable framework for understanding the various causes of seasonal extremes, and it is very well presented. I recommend the manuscript for publication with some minor adjustments. My principal concerns relate to the justification of the many choices which needed to be made in this analysis, these are detailed below.
Thank you, we changed the wording to "...our approach identifies 2-3 extreme seasons for each of winter, spring, summer and autumn, with strongly differing characteristics…".
2) L15: I think a justification of why 2 winter seasons were chosen for the in-depth case studies is needed here.
See answer to general remark (4) by reviewer 1.
3) L117: It is very nice to have these questions in the Introduction to frame the paper, but as far as I could see the synoptic systems of interest are pre-defined in the study (cyclones, blockings, and marine CAOs), so perhaps this question should be reframed to reflect this. This is indeed a good point. We rephrased question 3: "In which way do synoptic-scale weather systems such as cyclones, blocks and marine cold air outbreaks determine the sub-structure of extreme seasons?"

4) L131: What was the method(s) of interpolation?
This interpolation is done by the ECMWF software when downloading the ERA5 fields from the MARS archive.

5) L155
: What is the justification for choosing these regions?
As stated, a distinction between areas with differing sea-ice concentration is made, as surface heat fluxes and surface radiation are strongly dependent on the surface conditions. Further, we defined three different geographical regions, namely the Nordic Seas (NO), the Kara and Barents Seas (KB) and the remaining Arctic (AR). These are chosen based on the following main features: The NO region is the endpoint of the Atlantic storm track and important for deep water formation. The KB region has been strongly affected by changes in sea ice concentration and reacts very sensitively to atmospheric forcing. It is also a preferred region for atmospheric blocking and has its "own" storm track. Region AR is largely uncoupled from the mid-latitudes. Due to these different characteristics, it is useful to look at these regions separately when analysing the dynamical processes leading to Arctic extreme seasons. We added a sentence in this regard to the manuscript (L160ff.) to further justify the choice of our regions. 6) L161: Are results sensitive to the choice of definition of ice, mixed, and sea? Why were these thresholds chosen?
The results are sensitive to the choice of the SIC thresholds when defining ice, mixed and sea, because obviously the resulting regions get larger or smaller depending on how the thresholds are changed. For instance, if for ice, the threshold SIC clim was lowered from 0.9 to 0.8 then this would increase the size of the ice regions (and decrease the size of the mixed regions) and therefore the results for ice and mixed would be slightly less distinct. We decided to use relatively strict thresholds for ice and sea to ensure that these regions are indeed almost completely ice-covered and ice-free, respectively. 7) L174: Why choose just the first 2 PCs? This seems arbitrary, although I see later you mention that these explain a very large part of the overall variance.
See answer to first general comment of Reviewer 1. And yes, indeed the first two Principal Components explain usually 80-90% of the overall variance (in more detail: in 88% of the cases its >80% explained variance, in 53% of the cases they explain even >85% of the overall variance) and they are -for almost all regions and seasons -statistically distinct.
8) L178: Why are these rescaled by their respective SDs to give equal weight to each PC? Do you not wish to identify the extremeness of a season rather than the extremeness of a season with respect to these two PCs? (ref L114) If you don't do this rescaling do you still identify the same seasons as being extreme seasons?
We decided to use the scaled Euclidean distance (= Mahalanobis distance) in the PCA phase space to define our extreme seasons as with this approach, outliers in both, PC1 and PC2, are considered equally (without the rescaling, there would be more weight on the PC1 outliers). Thus, outliers in both PCs are treated similarly, independent of the individual variance explained by each PC.
We haven't tested the identification of extreme seasons without rescaling. Without rescaling, different, subjective thresholds for the definition of anomalous and extreme seasons would have to be chosen, which would hamper a direct comparison of the two methods. This is a very pragmatic and subjective choice. Results from a PCA might be less reliable for very small regions. With this threshold, each region comprises at least 40 model grid points.

30
In summary, this study shows that extreme seasonal conditions in the Arctic are spatially heterogeneous, related to different near-surface parameters, and caused by different synoptic-scale weather systems, potentially in combination with surface preconditioning due to anomalous ocean and sea ice conditions at the beginning of the season. The framework developed in this study and the insight gained from analyzing the ERA5 period will be beneficial for addressing the effects of global warming 35 on Arctic extreme seasons.

Introduction
Near-surface atmospheric conditions in the Arctic show a high variability on synoptic to inter-annual temporal scales, which is superimposed on a strong, long-term warming trend (e.g. Serreze and Barry, 2011;Cohen et al., 2014) :::::::::::::::::::::::::::::: (e.g., Serreze and Barry, 2011;Cohe Key drivers of variability on the synoptic to weekly time scale are interactions with the mid-latitudes for instance via air mass 40 exchanges (e.g. Woods et al., 2013;Laliberté and Kushner, 2014;Graversen and Burtu, 2016;Messori et al., 2018;Papritz and Dunn-Sigo air mass transformations within the Arctic (Ding et al., 2017;Pithan et al., 2018;Papritz, 2020). Both air mass exchanges and transformations are found to be related to synoptic weather systems. On longer time scales, in contrast, memory effects and feedback mechanisms such as the sea ice albedo feedback (Arrhenius, 1896;Curry et al., 1995), the water vapor and cloud feedbacks (Vavrus, 2004;Graversen and Wang, 2009;Boisvert et al., 2016), as well as the temperature feedback (Pithan and 45 Mauritsen, 2014) play an important role. Given this broad spectrum of processes, this leads to the question how variability on various temporal scales is inter-connected. In this study, we focus on the seasonal scale and it is our goal to analyze the role of intra-seasonal processes, including synoptic-scale weather systems, for the emergence of seasonal extremes in the Arctic.
The following paragraphs provide the relevant background on the key near-surface meteorological parameters in the Arctic and how they are interrelated. Furthermore, we discuss the role of different synoptic-scale weather systems for the variability of 50 these parameters and the occurrence of short-term extremes and seasonal anomalies in the Arctic.
Near-surface temperature, the components of the surface energy budget -including radiative and turbulent heat fluxes -as well as surface precipitation are especially important parameters linking the variability of the atmosphere with that of the ocean and the cryosphere. Large fluctuations in the surface energy budget, which themselves are closely linked to air temperature 55 fluctuations, contribute to the variability of sea ice (Stroeve et al., 2008;Olonscheck et al., 2019), the ocean mixed layer as well as open ocean convection (e.g. Marshall and Schott, 1999) ::::::::::::::::::::::::::: (e.g., Marshall and Schott, 1999). Radiative and sensible heat fluxes drive the variability of the surface energy budget components over sea ice (Lindsay, 1998), whereas over open ocean turbulent heat fluxes dominate (Segtnan et al., 2011). Precipitation variability influences snow cover, which is strongly linked to the albedo feedback, and it affects the freshwater balance of the Arctic Ocean and the Nordic Seas (Serreze and Francis, 60 2 2006; White et al., 2007), which jointly with turbulent heat fluxes impacts the thermohaline circulation (Dickson et al., 1996;Talley, 2008).
The three parameters -near-surface temperature, surface energy budget, and surface precipitation -do not vary independently from each other but they are interlinked. Thereby, the surface boundary conditions, i.e., sea ice vs. open ocean, strongly affect the type of linkages between parameters and of the :: as :::: well :: as : feedback processes due to vastly different heat capacities. On longer time scales, surface air temperature changes are largely influenced by variations in the sea surface temperature via surface sensible heat fluxes (Johannessen et al., 2016). In addition, incoming shortwave radiation is absorbed and can be 70 released to the atmosphere later. Over sea ice, in contrast, temperature is to a large degree determined by the surface energy balance, which includes radiative and turbulent heat fluxes, conductive heat fluxes across the ice and latent energy for freezing and melting (Serreze and Francis, 2006). In winter, when the incoming shortwave radiation is strongly reduced, the surface sensible heat flux and net surface longwave radiation mainly determine the surface energy balance in regions covered by sea ice (Ohmura, 2012). These considerations reveal that a meaningful identification of extreme seasons in terms of the surface 75 temperature, energy budget and precipitation parameters must take their co-variability and the underlying surface boundary conditions into account.
The role of synoptic-scale weather systems for inter-annual variability in the Arctic has been subject of multiple recent studies, which emphasized especially the importance of cyclones (Simmonds and Rudeva, 2012;Messori et al., 2018), blocking 80 anticyclones (Wernli and Papritz, 2018;Papritz, 2020), and Rossby wave breaking (Liu and Barnes, 2015). Air mass exchanges between the mid-latitudes and the Arctic region are often facilitated by cyclones, which, on one hand, transport warm and moist air to higher latitudes (Sorteberg and Walsh, 2008;Messori et al., 2018), causing there an increase in downward heat fluxes as well as the formation of clouds and precipitation. On the other hand, the advection of cold and dry air in the cyclones' cold sector enhances ocean evaporation and heat fluxes into the atmosphere. Additionally, extreme moisture transport into the Arctic 85 is often associated with events of Rossby wave breaking (Liu and Barnes, 2015), which can be strongly linked to the evolution of surface cyclones (Martius and Rivière, 2016). Air mass transformations within the Arctic can similarly result in anomalous conditions. Recent studies emphasized the importance of polar anticyclones and blocking events in the High Arctic, driving subsidence-induced adiabatic warming and thus leading to anomalies in surface temperature and net surface radiation, resulting in increased sea ice melting (Wernli and Papritz, 2018;Papritz, 2020). In winter, radiative heat loss under clear-sky conditions 90 can lead to extreme cold conditions, whereas cloud formation favors the trapping of longwave radiation, thus providing a positive warming feedback and causing an increase in surface temperature Woods and Caballero, 2016).
Similarly, a persistent and strong tropospheric polar vortex over the pole can isolate polar air masses and result in anomalously cold conditions due to enhanced radiative cooling (Messori et al., 2018;Papritz, 2020). Therefore, air mass transport and air mass transformation can significantly influence the Arctic surface energy balance. Whereas the modification of turbulent heat and caused significant sea ice melting in the Kara-Barents :::: Kara :::: and :::::: Barents : Seas . Binder et al. (2017) were able to show that several pathways of exceptional air mass transport caused this warm event. Another example is an extreme melt event on the Greenland ice shield in July 2012 (Nghiem et al., 2012), which was found to be related to a blocking anticyclone and associated anomalous long-range transport of warm and humid air masses from the South (Hermann et al., 2020). Such extreme weather events can have significant long-term effects, particularly due to their impact on sea surface  3. Which dynamical processes, in particular, which : In :::::: which :::: way ::: do synoptic-scale weather systems ::: such ::: as :::::::: cyclones, ::::: blocks :::: and :::::: marine :::: cold :: air :::::::: outbreaks : determine the sub-structure of extreme seasons?
4. What is the role of surface preconditioning , i.e., of early season anomalies of sea surface temperature and/or sea ice concentration for the formation of extreme seasons?
To address these research questions, a novel method will be introduced to determine the "unusualness" of a season, which 125 we define based on a combination of various surface parameters. Our study is organized as follows: Data and methods are described in Section 2. Section 3 presents an overview of the seasonal variability of surface temperature, surface precipitation, and of the surface energy budget components. In Section 4 we define anomalous and extreme seasons in the Arctic based on seasonal anomalies of these parameters, and analyze their substructure in distinct Arctic sub-regions. Detailed analyses of three Arctic extreme seasons and the involved atmospheric synoptic-scale processes are presented in Section 5, followed by the main 130 conclusions in Section 6.
190 P C1 and P C2 maximize the so-called "explained variance", which is the explained proportion of the total inter-seasonal variability in the six-dimensional phase space of the precursors.
To define extreme and anomalous seasons, P C1 and P C2 are first rescaled by their respective standard deviation (σ 1 and σ 2 ), such that outliers in both PCs are treated similarly independent of the variance explained by P C1 and P C2, thus providing 195 a measure for the unusualness of each season with respect to each of the principal components (from now on, we will refer to these rescaled components as P C1 and P C2). Then, the Euclidian distance in the reduced phase space spanned by the two rescaled components, the so-called "Mahalanobis distance" (d M ), is calculated as: This measure d M can now be used to quantify how strongly a particular season deviates from climatology, representing the 200 combination of the seasonal anomalies of the six variables. We therefore refer to d M as "anomaly magnitude" of a particular season. Seasons with d M ≥ 3 are defined as "extreme seasons", and seasons with 3 > d M ≥ 2 as "anomalous seasons". The phase space of the rescaled principal components can be illustrated using a biplot (Fig. 2), similar as in Graf et al. to the explained variance. If two vectors are approximately perpendicular, the precursors are uncorrelated. This interpretation of correlations is more precise, the higher the explained variance by PC1 and PC2 . The relative position of each season in the biplot (i.e., the scores) with respect to the precursor vectors indicates the contribution of the different precursor variables to the anomaly magnitude d M in the considered season. For instance, seasons with a positive T 2m anomaly are positioned in the direction of the T 2m vector and seasons with a negative T 2m anomaly in the opposite direction.

215
In the example given in Fig. 2, the variables T 2m and P show no correlation, whereas H S and H L are positively correlated and H S and P are strongly anti-correlated. Further, T 2m shows the largest contribution to the variance explained by PC1 and PC2 (mainly determining PC2) whereas H L , R S and R L mostly contribute to PC1. R S contributes the least to the explained variance. Two seasons with d M ≥ 3 are marked as extreme season 1 (ES1) and extreme season 2 (ES2). Their score vectors are 220 roughly orthogonal to each other, which indicates that a different combination of anomalies and thus different processes are decisive for explaining their large anomaly magnitudes. In this example, ES1 is mainly determined by a positive T 2m anomaly, while ES2 is an anomalously wet season with negative surface heat flux anomalies, as the respective precursor vectors are directed more or less directly towards (P ) respectively away (H L , H S ) from ES2.

Spatial and temporal variability of Arctic seasons 225
In order to characterize Arctic seasons in general, we first analyze the regional and temporal variability :::::::::: co-variability : of seasonal-mean anomalies of surface temperature (T 2m * ), precipitation (P * ) and surface energy balance (E S * ) :: in ::: the ::::: three :::::: regions, considering the varying surface conditions of the different sub-regions (Fig. 3). We are interested in correlations between the seasonal anomalies, how their magnitudes vary between the regions, and in aspects of the seasonal substructure (e.g., is an anomalously warm season constantly warm?).
In summer, the variability of the three analyzed parameters is smaller due to reduced meridional gradients of surface temperature and radiation causing smaller T 2m and E S fluctuations (Fig. 4b). Similar to winter, a large variability of T 2m occurs in the Kara-Barents :::: Kara ::: and ::::::: Barents : Seas and of P in the Nordic Seas. However, as the surface conditions between the sub-regions 265 become more homogeneous, the regions do not appear in distinct clusters as for winter with the exception of the sub-regions ARI and ARM, which cover most of the perennial sea ice and show, as in winter, only a small variability of the three parameters. It is further noteworthy that |T 2m * | :::::: |T 2m * | and |P * | are positively correlated, indicating higher variability in P in seasons with larger T 2m fluctuations.

300
The influence of the different surface conditions becomes apparent in particular for the seasonal substructure of E S anomalies. In regions with intermediate or high SIC clim , where surface heat fluxes are small and E S is mainly determined by net surface radiation, seasonal anomalies can continuously have the same sign, especially in summer, as shown by points near the diagonal in Fig. 6i and j. Over the open ocean, where surface heat fluxes are much more important, daily-mean E S values 305 fluctuate significantly around the climatology, which results in a large |E S * | but only a small E S * (Fig. 6l).
In KBM and KBS, a different distribution of E S anomalies occurs in DJF and JJA. Winters with a negative E S * , which is often caused by several episodes of cold air outbreaks (Papritz and Spengler, 2017), tend to show enhanced E S variability throughout the season (Fig. 5l) compared to winters with a positive E S * , where CAOs are less frequent. The opposite occurs 310 in summer, when periods of increased net surface radiation can cause a positive E S * and enhanced |E S * | compared to seasons with a negative E S * (Fig. 6k and l).
In winter, sub-regions over ice show a positive correlation between T 2m * and P * (Fig. 7a, d, g). This correlation is particularly 325 strong in the High Arctic, where precipitation events during winter are predominantly caused by synoptic weather systems that transport warm and moist air masses into the region (e.g. Webster et al., 2019;Papritz and Dunn-Sigouin, 2020) ::::::::::::::::::::::: (e.g., Webster et al., 2019; P T 2m * , P * and R L * mainly determine PC1 and thus the direction of maximum variance in the phase space spanned by all precursor variables in ice sub-regions. Surface sensible and latent heat flux anomalies are positively correlated and mostly uncorrelated with T 2m * and P * as they contribute mostly to PC2.

330
Similarly, sub-regions with intermediate sea ice concentration show a positive correlation of T 2m * and P * (Fig. 7b, e, h), although slightly weaker than over ice for regions KB and NO. Again, the heat fluxes are mostly uncorrelated with T 2m * and slightly negatively correlated with P * , particularly H L * . R L * is contributing less to the variance in mixed regions, which indicates a comparatively lower importance of radiation compared to heat fluxes for determining the seasonal variability.

335
Over the open ocean (Fig. 7c, f, i), a positive correlation between the heat flux anomalies and T 2m * can be observed, indicating increased surface fluxes from the ocean into the atmosphere during periods with anomalously cold temperatures. Unlike over ice, the maximum variance over open water is mainly determined by the surface heat fluxes. P * is mostly uncorrelated to the other variables and strongly related to PC2, reflecting that precipitation can occur in warm conditions (e.g., warm sector of 340 a cyclone) and in cold conditions (CAO).
Arctic summer seasonal variability is mainly determined by T 2m * , P * and R S * , whereby T 2m * and P * are mostly uncorrelated in all regions (Fig. 8). Whereas T 2m * shows only weak correlations with other parameters in general, P * is strongly anti-correlated with R S * in sub-regions NOS and ARS (Fig. 8f and i), most likely due to the presence of clouds during precipitation events. In sub-regions ARI and ARM (Fig. 8g and h), R L * additionally influences the seasonal variability and strongly correlates with T 2m * , again emphasizing the importance of clouds in this region.

DJF 2011/12
The winter of 2011/12 is classified as an anomalous season in KBI and KBM. In both sub-regions, this winter shows the largest positive T 2m * during the 39-year study period ( Fig. 7b and d). The time series in Fig. 9a shows that the daily-mean surface 395 temperature is continuously above the climatology (consistent with the fact that the dots in Fig. 5b and d are on the diagonal).
In KBI, T 2m * is the main contributor to this season's anomaly magnitude, supported by positive P * and R L * (Figs. 5b and f, and 7a). In KBM, positive T 2m * and H S * mainly determine the exceptional character of this winter (Figs. 5d and 7b), which also leads to one of the most positive E S * compared to all winters in the study period (Fig. 5l).
In DJF 2011/12, T 2m * is about + 6.6 K in KBI and + 4.7 K in KBM. In the whole region, during December, values are continuously around + 6 K above climatology, before approaching more normal :::::: average : levels at the beginning of January (Fig. 9a). The largest T 2m * values are reached in February. The SIC anomaly shows an opposite behavior and is continuously negative, reaching values close to climatology only at the beginning of the season and during the period with reduced T 2m * in January (Fig. 9c). Similarly to the other variables, we here calculate the SIC anomaly using a transient climatology, as this 405 effectively removes non-linear SIC trends in the Kara-Barents ::: Kara :::: and ::::::: Barents Seas (see Fig. S1c and d in the supplement).
Daily-mean E S values are strongly correlated with daily-mean T 2m , resulting in mostly positive E S * during the particularly warm episodes and shorter periods of negative E S * when T 2m * is reduced (Fig. 9b). The positive E S * is mainly due to a strongly positive H S * , i.e., strongly reduced heat fluxes into the atmosphere, favored by the warm surface temperatures and comparatively few CAOs (see next paragraph). During the period with the largest T 2m * in February, when the surface air tem- Several episodic precipitation events result in the strongly positive P * which often can be linked to the passage of a cyclone ( Fig. 10d, blue heatmap). Only very few episodes show P * values below climatology, e.g. at the beginning of February when the occurrence of a block causes dry conditions (Fig. 10d, red heatmap). The positive T 2m * results from several episodic warm events with a duration of ∼ 5-10 days (Fig. 10a), each deviating more than + 5 K from climatology. There are, however, also several periods that are notably colder than climatology, thus implying a small seasonal-mean anomaly. This ::::: These periods typ- Besides synoptic processes, also preconditioning potentially plays an important role for the occurrence of an extreme season, as we aim to discuss now. From Fig. 10c, it is obvious :: can ::: be :::: seen : that SIC in the Kara-Barents :::: Kara :::: and :::::: Barents : Seas was already exceptionally low at the start of the winter season, in fact, the sea ice extent on 01 December was the lowest 490 at :: on : this date for the entire study period. At the same time, the sea surface temperature (SST) shows a significantly positive anomaly of about + 1 K on average, which favors a delayed freeze-up in the region and at the same time also more intense upward sensible and latent heat fluxes. These initial surface conditions provide an important precondition for the strongly negative E S * , which itself is decisive for the anomaly magnitude of this winter. Analysing SIC and SST anomalies in the Kara-Barents :::: Kara ::: and ::::::: Barents : Seas during the previous seasons in 2016 shows that they developed since the previ-495 ous winter (SIC) or spring 2016 (SST, see Fig. 13b, :::::: which ::: will ::: be :::::::: discussed :: in ::::::::: section 5.3). At the end of 2015, an extreme warm event (e.g. Boisvert et al., 2016;Binder et al., 2017) ::::::::::::::::::::::::::::::::::::: (e.g., Boisvert et al., 2016;Binder et al., 2017) led to a significant thinning of the sea ice in the Kara-Barents :::: Kara :::: and :::::: Barents : Seas, causing an early start of the melt season in 2016 and subsequently increased SST values in MAM, coinciding with a positive T 2m * in the same region. The summer of 2016 does occur as an extreme season in sub-regions KBM and NOM (Fig. 8b and e) and as an anomalous season in KBS (Fig. 8c), mainly 500 due to a strong T 2m * of on average + 1.4 K in the Kara-Barents :::: Kara ::: and ::::::: Barents : Seas, which was facilitated by a reduction in total cloud cover and thus strongly enhanced R S . Together with the already existing positive SST anomaly this extremely warm summer led to record low SIC and ice-free conditions in the Barents Sea from July to September (Petty et al., 2018).
Strong blocking over large parts of the Arctic during autumn :::::: October :::: and :::::::: November : 2016 caused positive surface temperature anomalies across the whole Arctic region (Tyrlis et al., 2019) as well as strong positive E S anomalies, favoring the persistence 505 of the negative SIC and positive SST anomalies (Blunden and Arndt, 2017) until the beginning of DJF 2016/17.
In summary, the winter 2016/17 was extreme in the Kara and Barents Seas due to a combination of preconditioning and favourable synoptic conditions. Specifically, a combination of strongly positive SST * and negative SIC * at the beginning of the season, and a relatively large number of CAO events throughout the season, resulted in strongly negative surface heat flux 510 anomalies. Furthermore, an enhanced frequency of cyclones transporting warm and humid air masses into the region lead to a strongly enhanced P * . Comparing both anomalous winters in the Kara-Barents ::: Kara :::: and :::::: Barents : Seas, it becomes already evident from the PCA biplots ( Fig. 7a and b) that the processes leading to their respective anomaly magnitude are fundamentally different, as the vectors 515 pointing to the two seasons in the biplot are nearly orthogonal. The winter of 2011/12 is dominated by a continuous positive T 2m anomaly favored by a reduced frequency of CAO events, whereas in DJF 2016/17 the negative heat flux anomalies and exceptionally positive P * , enhanced by strongly reduced sea ice cover are most important. ::: We :::: have :::::: further :::: seen :: in :: the :::::::: previous :::::::::
6 Discussion and conclusions 590 In this study, Arctic winters (DJF) and summers (JJA) have been characterized based on seasonal anomalies of surface parameters including temperature, radiation, heat fluxes and precipitation for distinct regions considering varying surface conditions.
considering these spatial differences to study anomalous Arctic winter seasons. We further characterized Arctic seasons based on the seasonal substructure of surface temperature (T 2m ), precipitation (P ) and E S . Continuous seasonal anomalies, indicat-610 ing constantly anomalous conditions of the same sign throughout a whole season, can be observed for T 2m and E S except for the open ocean, where strong surface heat flux variability prevents continuous E S anomalies. Distinct outlier seasons can be observed featuring exceptional seasonal-mean anomalies in one or several parameter(s).
One of the main limitations of this study is the short time-period for which the ERA5 data is currently available. As our goal 665 is to study anomalous seasons, the number of suitable cases is strongly limited. Future analysis of large ensemble simulations of the CESM climate model will allow us to further statistically quantify and confirm the results of this study. The importance of long-term components such as the near-surface ocean processes leading to possible preconditioning of anomalous seasons have only been briefly considered in this study. Further analysis of anomalies in surface oceanic heat transport and its influence on sea ice formation and melt and sea surface temperatures will allow us to quantify the relative importance of short-term 670 atmospheric and long-term oceanic forcing in driving the processes leading to Arctic extreme seasons.
Code and data availability. ERA5 data can be downloaded from the Copernicus Climate Data Store (https://cds.climate.copernicus.eu/). The PIOMAS data set can be obtained from the Polar Science Center web page (http://psc.apl.uw.edu/research/projects/arctic-sea-ice-volumeanomaly/data/). Scripts used to produce the analyses and figures in this study are available on request from the authors.
Author contributions. KH performed most of the analyses, produced all figures and wrote the initial draft of the manuscript. All authors