Temporal evolution of relationships between temperature and circulation modes in five reanalyses

1 INTRODUCTION

Modes of atmospheric circulation variability (also called “teleconnections” in literature and referred to as “circulation modes” or only “modes” here for brevity) are composed of typically two or more action centres, between which sea-level pressure, geopotential heights, or other quantities describing atmospheric circulation are highly correlated (e.g., Barnston and Livezey, 1987). Such interconnections manifest in simultaneous strengthening or weakening of action centres, an example of which is the behaviour of two main centres of the North Atlantic Oscillation (NAO), the Azores high and Icelandic low: The pressure difference between them modulates the strength of westerlies blowing into Eurasia (e.g., Hurrell, 1995; Pokorná and Huth, 2015; Iles and Hegerl, 2017). Circulation modes affect various climate and environmental variables, such as temperature, precipitation, wind, snow cover, floods, and droughts, often far away from their action centres.

Relationships between circulation modes and climate variables (e.g., surface temperature, precipitation, wind) are not stable in time; they undergo periods of strengthening and weakening. This non-stationarity may be illustrated on the amplified impact (i.e., strengthened correlations) of NAO on temperature and precipitation over most of Eurasia roughly between 1960 and 1990 (Jung et al., 2003; Peterson et al., 2003; Polyakova et al., 2006; Beranová and Huth, 2007; Beranová and Huth, 2008; Sun et al., 2008; Vicente-Serrano and López-Moreno, 2008; Sun and Wang, 2012; Filippi et al., 2014; Xu et al., 2016; Zuo et al., 2016). Temporal non-stationarity of the relationship between NAO and climate variables was shown to be caused by a shift of its both action centres. In most of the second half of the 20th century, the centres of NAO are located more easterly, which causes stronger relationships with surface climate elements over most of Eurasia (Jung et al., 2003; Luo and Gong, 2006; Beranová and Huth, 2008; Sun et al., 2008; Vicente-Serrano and López-Moreno, 2008; Sun and Wang, 2012; Filippi et al., 2014). More eastward positions of the centres also give rise to an additional weaker positive centre over Central Asia, resulting in a strengthening effect of NAO in the Far East (Xu et al., 2016; Zuo et al., 2016).

The other dominant winter circulation mode in the Northern Hemisphere Extratropics is the Pacific/North American teleconnection pattern (PNA), the four centres of which are located over North America and eastern North Pacific (e.g., Leathers et al., 1991; Franzke et al., 2011). The structure of the PNA pattern changes mainly due to the eastward shift of the Aleutian low and the Canadian positive centre after 1980, which may have been invoked by changes in El Niño–Southern oscillation (He et al., 2013) or by a climate shift in the late 1970s (Lee et al., 2012). However, we are not aware of any study directly analysing the influence of the spatial changes of PNA on the stationarity of its relationships with climate variables. Furthermore, changes in relationships of other circulation modes in the northern Extratropics with climate variables have only been reported rarely (Krichak and Alpert, 2005; Beranová and Huth, 2008).

Circulation modes are frequently identified by principal component analysis (PCA) of geopotential heights in mid- or lower troposphere with an atmospheric reanalysis as a data source. Depending on the type of assimilated data, reanalyses can be divided into two groups (Fujiwara et al., 2017). Full-input reanalyses assimilate all available data including surface, satellite, and radiosonde data. Surface-input reanalyses are based solely on surface observations such as pressure, surface winds, and sea-ice distribution. Because the surface-input reanalyses utilize only a limited amount of surface observations, they can extend farther into the past and usually span more than century-long period (20CRv3 starts in 1836). On the other hand, free atmosphere is completely calculated by the model in the surface-input reanalyses, which may bring about inhomogeneities and imperfections or even errors in the data set.

Studies comparing reanalyses indicate that the agreement among full-input reanalyses (JRA-55, NCEP-1) is generally very good, while the reanalysis most different from the others is surface-input reanalyses 20CRv2(c), in which biases were found to be located over central and southern Eurasia and northern to north-eastern Europe (Wang et al., 2016; Stryhal and Huth, 2017; Rohrer et al., 2018; Hynčica and Huth, 2020b). These biases affect the shape and position of some circulation modes, most strongly Eurasian mode type 1 (EU1) in winter (Hynčica and Huth, 2020b). The newly issued version of the 20CR reanalysis, 20CRv3, is not expected to suffer from these errors any longer (Slivinski et al., 2019; Slivinski et al., 2021). The other widely used surface-input reanalysis, ERA-20C, shares a larger congruity with full-input reanalyses (Stryhal and Huth, 2017; Hynčica and Huth, 2020b).

Although the papers mentioned above compare various climate variables between reanalyses, no study has compared temporal variations in their relationships with circulation modes yet. The goal of this paper is therefore to evaluate differences in relationships between circulation modes and temperature in five reanalyses in winter. Our aim is the comparison and identification of possible causes of differences. We do not intend to describe mechanisms causing the temporal non-stationarity of relationships with circulation modes.

2 DATA AND METHODS

We compare relationships of circulation modes with temperature from January 1958 to December 2010 in winter over the Northern Hemisphere Extratropics (delimited by 20°N). We analyse the longest period possible with respect to the temporal coverage of all reanalyses.

The analysis is conducted for monthly mean 500 hPa geopotential height anomalies on a 5 × 5 regular grid in three surface-input reanalyses (ERA-20C, 20CRv2c, and 20CRv3) and two full-input reanalyses (JRA-55 and NCEP-1; for more details, see Table 1). The original resolution of 20CRv2c is 2 × 2; therefore, it was interpolated to the requested resolution by bicubic splines. Furthermore, the convergence of meridians towards the pole, which may bring biases into the analysis (Huth, 2006), is accounted for by forming a quasi-equal-area grid, that is, by excluding some gridpoints, so that the mean area of a gridbox is kept approximately the same at all latitudes.

Circulation modes are calculated by rotated PCA in an S-mode (spatial locations are organized in rows and time realizations in columns of the data matrix) from the correlation matrix. For each principal component (PC), PCA determines a loading (a spatial representation, i.e., correlation with the original data field) and a score (time series of the intensity of the mode). The PCs are subject to orthogonal Varimax rotation, which improves the interpretability of components; that is, loadings are then more regionalized (Barnston and Livezey, 1987; O'Lenic and Livezey, 1988; Lian and Chen, 2012). There are many options of how to determine the number of PCs to rotate (e.g., Richman, 1986); here, we employ the approach based on a scree plot. The scree plot is a graph showing the dependence of explained variance on the order of PCs, which are arranged on the x-axis in a descending order (O'Lenic and Livezey, 1988). There are usually drops on the scree plot; the rightmost drop typically indicates the number of components to rotate. However, the drop may not always represent the optimal value for rotation because some circulation modes might be located behind the drop and therefore would be lacking in the analysis; thus, it is sometimes better to rotate one additional component (Hynčica and Huth, 2020b). For example, the drop is localized to the 10th component in all reanalyses but 20CRv3, in which it occurs at the ninth component. Nevertheless, we prefer to keep 10 rotated components also in 20CRv3 because the additional (10th) mode is recognized in all other reanalyses.

Surface temperature data are obtained from the Climatic Research Unit gridded Time Series Dataset v 4.03 (Harris et al., 2020, the CRU data set in further text) in its original 0.5 × 0.5 resolution. Monthly temperature anomaly from the long-term average is computed at each gridpoint. Since relationships with atmospheric circulation are mostly similar at stations and their nearest gridpoints in the CRU data set (Hynčica and Huth, 2020a), we use the gridded data set here because it provides data in a regularly distributed form, which is more suitable for spatial analysis.

Temporal evolution of relationships is evaluated by running correlations between all circulation modes in all reanalyses and temperature anomalies at all gridpoints. A moving 45 months (i.e., 15 winters) window (running period) with 1-month shift is employed; the value of running correlation is labelled to the central month of the window. Hence, time series of running correlations (116 values starting in 1965 and ending in 2003) between each circulation mode in five reanalyses and temperature anomaly at every gridpoint is created. Differences between running correlations for all pairs of reanalyses are quantified by the absolute value of the average difference between time series (mean absolute differences; MAD) for each circulation mode and each gridpoint in the CRU data set.

3 GENERAL EVALUATION

Ten circulation modes explaining around 70% of total variance are identified in all reanalyses. Circulation modes in JRA-55 are displayed with their respective abbreviations in Figure 1 (circulation modes in other reanalyses are shown in Figures S1–S4). The appearance of nine circulation modes is largely similar to other hemispheric studies (Barnston and Livezey, 1987; Clinet and Martin, 1992; Huth et al., 2006; Hynčica and Huth, 2020b). However, one of the modes, denoted as “Mode5” here, does not occur in most analyses. It is only listed in Huth et al. (2006) as “Mediterranean mode,” which appears solely in maxima of solar activity when it seems to represent detached eastern parts of NAO. On the other hand, Mode5 is attached to EU2 in some reanalyses for a different period 1957–2002 (Hynčica and Huth, 2020b). The realism of the mode is supported by autocorrelation maps for its centres, which closely resemble the loading pattern (not shown). Interestingly, EU2 is not identified as a separate mode in 20CRv3 even when various numbers of components are rotated.

Figure 2 shows running correlations with selected circulation modes at a gridpoint in southern Greenland as an example. One can see a considerable temporal variability of running correlations for all modes in all reanalyses. Even this single example shows that 20CRv2c is outlying from other reanalyses, which is particularly evident for NAO and EU1. The relative behaviour of the time courses of correlations is varied: At this gridpoint, the differences between reanalyses tend to be smaller for some modes (TNH), time courses may tend to be parallel to each other, that is, differences between reanalyses tend to be constant over time (NAO), while for some other modes, differences tend to vary over time for most pairs of reanalyses (Mode5, EU1). Differences between running correlations at this gridpoint are summarized in Table 2. Smaller MADs correspond to similar time courses of running correlations in a pair of reanalyses (e.g., all the four circulation modes between JRA-55 and NCEP-1). One can see that differences are largest for 20CRv2c.

MADs are then aggregated over all gridpoints and displayed in boxplots in Figure 3 for each circulation mode (aggregating also over all reanalysis pairs) and for each reanalysis pair (aggregating also over all modes). Larger values of MADs indicate larger differences of time series of running correlations between reanalyses. The largest differences occur for Mode5, EU1, and NAO, while the time series of relationships are most similar for PNA and EP (Figure 3, top). If we focus on pairs of reanalyses, both 20CR reanalyses differ the most from the rest of reanalyses, and this is particularly so for the older version (20CRv2c), making it the most outlying reanalysis. NCEP-1, JRA-55, and ERA-20C exhibit a large congruity with each other (Figure 3, bottom). A more detailed insight is offered by Figure 4, which displays the median of MADs for each circulation mode and pair of reanalyses. It is evident that larger discrepancies in running correlations occur mainly for Mode5 in all reanalyses, NAO in 20CRv2c, and EU1 in both 20CR reanalyses.

We also display maps of MADs for two modes (NA and EA) and three pairs of reanalyses (Figure 5) as an example. Maps are organized so that MAD increases from top to bottom. The magnitude of MAD is primarily determined by the spatial similarity of circulation modes between reanalyses: a large spatial resemblance of the modes leads to a low or even negligible MAD (Figure 5, top row). In contrast, a spatial shift of the circulation mode between reanalyses leads to regionally increased MADs, as exemplified for NA in 20CRv3 and EA in 20CRv2c (Figure 5, bottom row). The relative shift of both modes (see Figures S1 and S2) subsequently results in a regionally enhanced dissimilarity of relationships and hence larger MADs. Additionally, correlations of scores (displayed in brackets on the top of each map in Figure 5) tend to relate proportionally to the similarity of relationships. This is, however, not always true as illustrated for the EA mode between 20CRv2c and NCEP-1 (Figure 5, bottom row, right): Large MADs appear despite high correlation of scores (0.95). The scores of EA in 20CRv2c are similar to other reanalyses (correlations exceeding 0.93; Table 3) although its spatial pattern is shifted (Figure S1). A comparison of medians of MADs (Figure 4) and correlations of scores for all reanalyses (Table 3) indicates that this applies also to NAO in 20CRv2c and NA in 20CRv3.

4 ANALYSIS OF DIFFERENCES IN 20CR REANALYSES

The comparison of reanalyses reveals that differences of relationships of temperature with circulation modes are largest for NAO and EU1 in 20CRv2c, for EU1 in 20CRv3, and for Mode5 in all reanalyses. Mode5 has a zonal orientation with a circumglobal belt located in low latitudes with little pronounced anomaly circulation related to it (see Figure 1); such circulation modes tend to have a lower agreement among reanalyses (this also applies, e.g., to the subtropical zonal pattern in spring, summer, and autumn; Hynčica and Huth, 2020b). Differences related to Mode5 are likely to result in only small discrepancies in correlations with temperature in mid-latitudes since the correlations themselves are small there (not shown); a detailed analysis of Mode5 is, therefore, of little potential interest. That is why we mainly focus on the 20CRv2c reanalysis in the investigation and interpretation of specific differences.

4.1 Example 1: EU1 in 20CRv2c and NCEP-1

Now we look at differences in running correlations between reanalyses in more detail. We take mode EU1 and the 20CRv2c and NCEP-1 reanalyses as the first example. Results are representative for differences between the 20CR family and the rest of reanalyses; that is, they are qualitatively similar if 20CRv3 is taken instead of 20CRv2c and JRA-55 or ERA-20C instead of NCEP-1.

The positions of both major action centres of EU1 (the positive centre in northern Europe and the negative centre over Central Asia) differ between the two reanalyses (Figure 6, top). This leads to a different evolution of running correlations in two compact areas: Central, Northern, and north-eastern Europe, and a belt across north-eastern Africa and south-western Asia (marked as A and B in Figure 6, respectively). A more westward position of the northern centre in 20CRv2c leads to a potential for a stronger intrusion of cold air into Europe when EU1 is in its positive phase, implying more negative correlations. The negative centre over Asia is weaker and smaller in 20CRv2c, which leads to weaker anomaly circulation and correlations closer to zero over South-west Asia and North-east Africa (Figure 6, middle). A notable feature is that the magnitude of differences between correlations in the two reanalyses varies in time, peaking around 1985 in both areas (Figure 6, bottom).

In order to identify causes of the time variation of differences between reanalyses, composite maps of EU1 are calculated for all running periods in both reanalyses. For each running period, the composite map contains differences of 500 hPa anomalies between months with the score of EU1 higher than +1 and months with the score lower than −1. We define the position of action centres in the composite maps by the gridpoints with the largest or lowest anomaly in two longitudinally limited domains (75°W–40°E for the European centre and 42.5°W–145°E for the Asian centre). Then, the time series of strength, latitude, and longitude of both centres are compared between reanalyses.

The composites and composite differences in Figure 7a show that the two centres of EU1 move differently in the two reanalyses and that their relative position varies in time. The European centre is located more westerly in 20CRv2c in almost all running periods (Figure 7b, top left). It starts moving westward earlier (in 1979) in 20CRv2c than in NCEP-1 (in 1986), resulting in the largest longitudinal difference to occur in 1985 (marked as T2). Later, the centre moves back eastwards, again earlier in 20CRv2c, resulting in a diminished longitudinal difference between the reanalyses near the end of the analysed period (running period T3). The northward move of the centre is similar in both reanalyses (Figure 7b, middle left); its overall strengthening is less synchronized between the reanalyses, the centre being stronger in 20CRv2c from 1983 to 1995 (Figure 7b, bottom left). The similarity of the position of the European centre and its spatial structure between the reanalyses near the beginning and end of the analysed period (including running periods T1 and T3) are the reasons for small differences in correlations of EU1 with temperature in Europe (Area A in Figure 6). On the other hand, the potential for an inflow of cold air from north-west and north into most of Europe under the positive phase of EU1 is particularly effective in the mid-1980s (Period T2) in 20CRv2c, thanks to the European centre being stronger and located more westward. This is reflected by more negative correlations in 20CRv2c compared to NCEP-1.

The differential position and intensity are not sufficient for the explanation of different running correlations near the Asian centre. Although the correlations in the area affected by it (north-eastern Africa and south-western Asia, B in Figure 6) differ between the two reanalyses in Running Period T2 as much as in Area A, the position and intensity of the Asian centre are fairly similar in both reanalyses (Figure 7b, right). The reason for the difference in correlations appears to consist in the shape of the Asian centre: In 20CRv2c, it lacks the south-westward extension across Central Asia towards Iran, which is rather strong in NCEP-1. As a consequence, the positive phase of EU1 supports the intrusion of cold air from the north into Area B in NCEP-1, whereas this effect is absent in 20CRv2c. This explains negative correlations in Area B in period T2 in NCEP-1 and less negative/more positive correlations in 20CRv2c.

4.2 Example 2: NAO in 20CRv2c and NCEP-1

Spatial representation of NAO is also slightly different in the two reanalyses: Both the major centres of NAO are shifted westward to south-westward, and the northern negative centre is considerably weaker in 20CRv2c (Figure 8, top). Such spatial setting consequently induces weaker (less negative and less positive, respectively) running correlations (and larger MADs) in 20CRv2c over the southern half of Greenland and in a long belt over northern Eurasia (Areas A and B, respectively, in Figure 8, middle).

Over northern Eurasia, the correlations are closest in the two reanalyses at the beginning of the studied period, followed by the increase of differences until the late 1970s. On the contrary, differences culminate in the 1990s over southern Greenland (Figure 8, bottom). Composite maps are calculated in the same way as for EU1; positions of the centres are detected between 100°W and 70°E. Figure 9a shows composite maps in the three selected periods: T1 is a period with small differences in running correlations, particularly in northern Eurasia, while the other two periods correspond to increased differences over northern Eurasia (T2) and southern Greenland (T3) (Figure 8). Time series of the position and strength of the NAO centres in both reanalyses are shown in Figure 9b. Note that composites could not be calculated between 1997 and 2002 because of missing negative phase: There are no months with the score of NAO below −1 in the corresponding periods.

The positive relationship between temperature and NAO over large parts of Eurasia is determined by the intensity of westerlies, which is modified by the position, strength, and shape of the NAO centres: A strong positive NAO induces strong westerlies, which bring warm air into most of Europe, which results in positive correlations of the intensity of NAO with temperature. Over northern Eurasia, running correlations started to differ after 1965 with the differences peaking before 1980 (T2; Figure 8, middle). In spite of the similar position and strength of the southern positive centre in both reanalyses in T2 (Figure 9a, middle; Figure 9b, right), its spatial appearance differs between the reanalyses: the extension towards Eurasia is weaker and less pronounced in 20CRv2c. The centre spreads far into northern Siberia in NCEP-1, which together with a more poleward located northern centre account for stronger westerlies penetrating deeper into Eurasia, and therefore stronger correlations compared to 20CRv2c.

In T1, a closer agreement in the position and intensity of both centres and consequent lower composite differences over Atlantic and Europe (Figure 9a, top) result in similar relationships over Northern Eurasia. Despite the centres being fairly similar also after 1990 (T3, Figure 9a, bottom), relationships differ (Figure 8). The reason is in the positive centre over Northern Eurasia (Figure 9a, bottom), which is responsible for weaker relationships with NAO due to stronger meridional circulation. This centre is stronger and more extensive in 20CRv2, which elicits stronger anomaly circulation and hence more pronounced decrease in correlations.

Relationships between NAO and temperature in Greenland are driven by the northern centre, which controls advection of cold northerly/north-easterly air into the region. The centre is permanently placed more southward in 20CRv2c (Figure 9b, left), which results in weaker advection of cold air into southern Greenland under the positive phase of NAO and subsequent less negative correlations. The weakening of the northern centre in 20CRv2c during the late 1970s and early 1980s (Figure 9b, left) was not accompanied by a change in relationships with temperature. However, correlations became less negative in 20CRv2c after 1990 (Figure 8a, bottom) without a prominent change in the position and intensity of the centre. Differences in the shape of the centre are a probable cause of differences in correlations.

5 CONCLUSIONS

Temporal evolution of relationships between circulation modes and temperature in the northern Extratropics in winter was compared between five reanalyses. Relationships are defined by running correlations calculated for 45-month windows (corresponding to 15 winter seasons) between the intensity of circulation modes (provided by rotated PCA) and temperature anomalies from 1958 to 2010. Running correlations with all circulation modes are in high accordance with the full-input reanalyses (NCEP-1 and JRA-55) and ERA-20C; this is in accord with Hynčica and Huth (2020b) where a high congruity of spatial patterns of circulation modes between these reanalyses is reported especially for winter. Both versions of the 20CR reanalysis differ from other reanalyses more substantially, which is particularly so for the older version, 20CRv2c. The modes located over Eurasia and the North Atlantic have different positions and differ in correlations or even are not detected at all. This concerns EU1 (in both 20CR reanalyses), NAO (in 20CRv2c), EU2 (not identified in 20CRv3), and to a lesser extent EA in 20CRv2c and NA in 20CRv3. The other circulation modes and related correlations with temperature in 20CRv2c and 20CRv3 are largely consistent with other reanalyses; this applies mainly to those over the Pacific and North America (PNA, WPO, TNH, and EP). Overall, 20CRv3 is closer to other reanalyses than 20CRv2c, suggesting that the development of the 20CR reanalysis is successful in the reduction of its errors and biases. The large congruity of the other surface-input reanalysis, ERA-20C, with the full-input reanalyses, indicates that even the surface-input reanalyses have the potential to accurately reproduce atmospheric circulation in mid-troposphere, at least in periods when surface data are sufficiently dense and reliable. This concerns mainly the second half of the 20th century when the amount of the input data substantially increased, while the lack of data in the preceding period is responsible for a lower reliability of reanalyses.

The differences between reanalyses in correlations of the modes with temperature, their spatial distribution, and development in time are attributed to differences in the position and strength of the action centres of the modes, uncovered by the composite analysis. The different patterns of the modes induce different anomaly circulations, which in turn leads to a different temperature response to the intensity of the mode and different correlations of the intensity of the mode with temperature.

AUTHOR CONTRIBUTIONS

Martin Hynčica: Conceptualization; data curation; formal analysis; investigation; methodology; software; validation; visualization; writing – original draft. Radan Huth: Conceptualization; formal analysis; funding acquisition; methodology; project administration; resources; supervision; writing – review and editing.