Reanalysis representation of low-level winds in the Antarctic near-coastal region

,


15
Easterly tch: surface winds prevail over most of coastal Antarctica and extend up to about 2 km above surface level. Easterly wind stress is in turn an important driver of the westward Antarctic Coastal Current, a circulation coupled in many regions to the Antarctic Slope Front (ASF) between shelf waters and warmer circumpolar deep water (Thompson et al., 2018). The Southern Ocean meridional overturning circulation may respond more to changes in Antarctic coastal winds (or 'polar easterlies') than to the westerlies further north (Stewart and Thompson, 2012). It has been shown that modifications to easterly wind stress 20 and Ekman pumping near the coast could enhance subsurface advective heat fluxes across the ASF, contributing to ice shelf basal melt (Spence et al., 2014;Goddard et al., 2017;Stewart et al., 2019) tch: . Antarctic coastal winds also have a large noneasterly component, especially on short timescales. Meridional winds tch: , while coastal katabatic winds could act as a control on 1 ice shelf stability in some regions (Hazel and Stewart, 2020) tch: . Coastal winds also modifyand have a major impact on sea ice concentrations, for example off the Ross Ice Shelf (Kurtz and Bromwich, 1985;Petrelli et al., 2008;Mathiot et al., 2012;25 Silvano et al., 2020) tch: , and in other prominent regions of sea ice production (Wang et al., 2021). Given the sparsity of nearby observations, reanalysis datasets are critical for characterising Antarctic coastal winds, and often provide boundary conditions for atmospheric and ocean models. As an observational benchmark they are also important for evaluation of coupled models used for future projections.
In part due to the paucity of assimilated observations, historical trends of tch: the coastal easterliescoastal winds from reanalysis are 30 uncertain and differ among datasets, though all indicate an increase in tch: their seasonalitythe seasonality of the easterlies over the last few decades (Hazel and Stewart, 2019). Over Antarctica itself, the magnitude of the late 20 th century trend in sb: surface wind near-surface winds differs considerably in its pattern and magnitude between reanalyses, especially in the coastal region (Dong et al., 2020). On average, reanalyses exhibit higher wind speed correlation coefficients with surface station measurements in summer compared to winter and over inland regions compared to the coast (Tetzner et al., 2019;Dong et al., 2020). The realism 35 of tropospheric wind profiles is improved in the more recent ERA5 reanalysis with respect to the earlier ERA-Interim at coastal stations, but large deficiencies remain (Vignon et al., 2019). Reanalysis evaluations have tended to focus on the onshore coastal region, in part due to the limited availability of satellite observations near the coast (Bourassa et al., 2019), but, since offshore winds play a key role in interactions between the atmosphere, ocean and cryosphere, an evaluation of performance against observations from the marine coastal sector is needed. 40 Alongside the lack of evaluation offshore, relatively little attention has been given to reanalysis representation of wind variability and extremes. Jones et al. (2016) show how considerable underestimation of low-level wind speeds occurs in the Amundsen Sea Embayment when a 'low-level jet' is present, reminiscent of underestimated elevated wind speeds in regions of complex orography in the Arctic (e.g., Moore et al., 2016). Extreme coastal winds are important for rapid sea ice loss (Jena et al., 2022) and calving events (Francis et al., 2020) so a characterisation of reanalysis performance during high-wind states is 45 needed. However, Antarctic coastal winds are influenced by a range of drivers, each of which presents a different set of challenges for reanalysis representation. Although mean coastal wind speeds peak onshore, the variability on 12-hourly timescales is highest just offshore where the interaction between directionally-constant continental winds and highly variable synoptic flows becomes important ( Figure 1). One important driver is katabatic forcing, which sustains shallow terrain-following drainage flow towards the coast (Ball, 1960;Parish and Bromwich, 1987) but its tch: offshore extent is a source of uncertainty as the flow 50 often stops abruptly at the coastal margin (e.g., Yu and Cai, 2006;Tomikawa et al., 2015;Vignon et al., 2020) and its behaviour in models is sensitive to the representation of the atmospheric boundary layer (King et al., 2001;Parish and Cassano, 2003;Orr et al., 2014). tch: As a result, the indirect effect of katabatic forcing on offshore flow, for example due to momentum advection, is uncertain. Representation of orography is important as long-wave cooling over the steep coastal slopes sets up both katabatic flows near the surface and a deeper layer of terrain-following isentropes above it which explains the easterly 'low-level jet' 55 structure (Fulton et al., 2017). Orography is also critical for blocking and barrier winds induced by steep Antarctic coastal slopes Cassano, 2001, 2003;Orr et al., 2014;Weber et al., 2016;Yamada and Hirasawa, 2018). Other drivers include the large-scale pressure gradient force (van den Broeke and van Lipzig, 2003) and, on shorter timescales, synoptic storms.
In this paper, we aim to provide both a representative overview of reanalysis performance in the coastal easterly sector but 60 also to test whether that performance is sensitive to short-term variations in the large-scale pressure conditions driving the coastal winds. By comparing reanalysis performance across driving states and geographic locations, we aim to shed light on possible sources of error and offer a process-oriented perspective to help select the most appropriate reanalysis to use when characterising the Antarctic coastal easterlies. tch: Throughout this paper, we refer to 'synoptic' and 'katabatic' forcing to characterise short-term variability. By 'synoptic' we are referring to variations in flow linked to large-scale pressure gradients (order 65 of 1000 km) set up over a timescale of days, for example due to the passage of low pressure systems. 'Katabatic' forcing is instead associated with near-surface thermodynamic processes driving downslope flow, with uncertain variability through time.
The research in this paper is guided by three questions: 1. How well do reanalyses represent the spatial patterns and short-term variability of sb: surface wind near-surface winds within the Antarctic coastal easterly sector?
To what extent is reanalysis representation of Antarctic coastal winds sensitive to variations in the driving flow regime?
3. Which reanalysis most accurately represents tch: the Antarctic coastal easterliesAntarctic coastal winds?
Our analysis has two components; firstly, to quantify reanalysis performance across the entire coastal region we compare with observations from the EUMETSAT MetOP-ASCAT sensor, which has near-coastal data available during austral summer.
Next, to evaluate reanalysis representation of variability on short timescales, a self-organising map (SOM) methodology is 75 used to composite surface and upper air wind data available from four coastal surface stations which have high-resolution radiosonde data compiled by Vignon et al. (2019).
Section 2 details the datasets used as well as the collocation and SOM approach. Results are presented in Section 3, including an evaluation against ASCAT (3.1) and a SOM-based evaluation against station measurements (3.2). Results are discussed and an overall assessment of reanalysis reliability is given in Section 4 with conclusions in Section 5.

Reanalysis datasets
For this analysis, we evaluate ERA5, MERRA-2 and JRA-55, which belong to the latest generation of reanalyses which have been widely adopted for studies of the Antarctic circulation as well as for quantifying historical trends and evaluating climate models. tch: Near-surface winds are evaluated against in-situ and scatterometer-derived winds for the period January 2010 to 85 December 2017. tch: For the comparison with scatterometer winds the reanalysis neutral 10 m winds are used. This is because the same assumption of a neutral profile is made when deriving ASCAT 10 m winds from surface stress (the feature directly observed by scatterometers) so, as in the ECMWF assimilation scheme Hersbach (2010), we consider them to be more comparable.

90
ERA5 is the fifth generation in the European Centre for Medium-range Weather Forecasts (ECMWF) series of reanalyses.
A full technical description of the setup is given by Hersbach et al. (2020). The dataset is derived using a 2016 version of the operational ECMWF Integrated Forecasting System (IFS) model (cycle 41r2) and a hybrid incremental 4D-Var data assimilation methodology. ERA5 is available on a regular 0.25 • grid at 31 km horizontal grid spacing and on 137 vertical levels. Both surface-level diagnostics and data on pressure levels are available hourly. A final release of ERA5 is publicly 95 available for 1959 onwards. ERA5 sb: surface wind near-surface winds are a diagnostic derived in IFS output by scaling the winds at the 40 m height level to 10 m assuming a roughness length for short grass (0.03 m) when the surface roughness is less than or equal to 0.03 m and otherwise using the tile roughness length. For our comparison with ASCAT ocean winds, the ERA5 10 m neutral wind product is used instead, which is derived from surface stress using the tile roughness length and assuming neutral stability. The Modern-Era Retrospective Analysis for Research and Applications, version 2 (MERRA-2) has been developed by NASA's Global Modeling and Assimilation Office (GMAO) as an intermediate reanalysis intended to incorporate updates to GMAO modelling and data assimilation into the MERRA framework (Gelaro et al., 2017). MERRA-2 is based on the Goddard Earth Observing System-5 (GEOS-5) forecast model and is available from 1979 on an approximately 0.5×0.625 • grid (50 km grid 105 spacing) with 72 vertical levels. Surface fields (including 10 m wind speed) are available hourly, with 3-hourly vertical fields on pressure levels. Wind speeds at 10 m are derived in MERRA-2 by interpolating from the lowest model level using Monin-Obukhov similarity theory, thereby accounting for the effects of varying near-surface atmospheric stability (Torralba et al., 2017). These stability-dependent winds are compared with station measurements. For comparison with ASCAT, we calculated 10 m neutral wind speeds using the hourly output of friction velocity (u * ) and roughness length (z 0m ) from the MERRA2 flux 110 diagnostics (GMAO, 2015) following: where k=0.4 is the von Kármán constant.

JRA-55
The Japanese 55-year Reanalysis (JRA-55) developed by the Japan Meteorological Agency (JMA) covers a period from 1958 115 onwards and is an update to JRA-25, incorporating novel observations and a 4D-Var data assimilation system (Kobayashi et al., 2015). The JMA global spectral model (GSM) is used as the underlying forecast model. The JRA-55 grid corresponds to approximately 55 km horizontal grid spacing, with 60 vertical levels. Surface fields are available on six-hourly timesteps, so compared to ERA5 and MERRA-2 the comparison with relatively infrequent scatterometer observations is more limited (Section 2.2.1). JRA-55 10 m winds are interpolated from the lowest model level as in MERRA-2, but unlike the stability-120 dependent ERA5 and MERRA-2, sb: surface wind near-surface winds are calculated assuming neutral stability in the surface layer (Torralba et al., 2017). Scatterometer (ASCAT, version 2.1) is a C-band scatterometer used to measure surface radar backscatter along two swaths each ∼500 km wide separated by a 360 km gap processed to provide wind vectors for a grid spacing of about 25 km. The data contain a rain flag, since rain can produce a positive bias at low wind speeds due to backscatter from rain drops and negative 130 bias at high wind speeds due to atmospheric attenuation of the signal. Undetected sea ice can limit the quality of scatterometer data as well as wind speeds higher than 35 ms −1 (Ricciardulli and Wentz, 2016). Rivas and Stoffelen (2019) report that the magnitude of ASCAT wind component errors on the scatterometer measurement globally is ∼0.7 ms −1 with negligible bias.
tch: We use ASCAT to evaluate reanalysis near-surface winds offshore, but it is important to note that ASCAT wind observations are assimilated into all three reanalyses used in this paper from 2008. tch: A close correspondence between ASCAT and the 135 reanalyses is expected, therefore. Remaining differences could be due to large biases in the first guess model, inadequacies in the assimilation system, scale mismatches and inclusion of observations in our results which were not assimilated by the respective reanalyses. We do not intend to provide an independent evaluation of skill using ASCAT, but instead to compare between reanalyses, and to highlight aspects of the observed wind field and variability which are not captured by the reanalyses in spite of the observational constraints applied.  Analysis is primarily constrained to austral summer due to the seasonal expansion of sea ice which prevents wind speed observations near the coast. A breakdown of the results by season is given in tch: AppendixSection S2 tch: (supplement).  (Figure 1). 10 m wind speed measurements from surface stations are withheld from assimilation into JRA-55, MERRA-2 and ERA5 (Kobayashi et al., 2015;Gelaro et al., 2017;Hersbach et al., 2020), but the local wind field is not entirely independent from observations due to the assimilation of upper air wind, pressure and temperature profiles. Surface meteorological observations were obtained via the SCAR READER database (Turner et al., 2004). Neumayer, Mawson and DDU were selected for this analysis to evenly sample the East Antarctic coastline as well as due to their high representativeness 160 of the coastal lower troposphere (see Vignon et al. (2019), Figure 11). Casey, on the other hand, is less representative but is situated near complex orography which is represented differently in the three reanalyses (see Section 2.3.1). Long-term vertical profile data from West Antarctica are currently lacking but would be highly valuable for an evaluation of reanalysis performance of this kind, for example under various configurations of the Amundsen Sea Low. outdoor temperatures prior to launch) and to allow the balloon to reach ambient flow velocity.

Collocation of reanalyses and station measurements
Comparison of gridded datasets with point observations is not straightforward, especially in regions of complex orography or where steep slopes generate an offset between the actual station elevation and that at the station location in the model (Dong et al., 2020). In this evaluation we compare station observations with the nearest reanalysis grid point for both the surface and source of topographically induced gravity waves with highly localised effects (Murphy and Simmonds, 1993;Turner et al., 2001;Adams, 2004). All reanalyses include some representation of the Law Dome but in ERA5 the peak rises to 1,259 m compared to 964 m in MERRA-2 and 952 m in JRA-55.
Orographic slopes from the plateau to the coast differ considerably between reanalyses ( Figure 2, rightmost column). All three datasets have orography that is noticeably smoothed compared to the 1 km Radarsat Antarctic Mapping Project Digital

190
Elevation Model (DEM) version 2 (RAMP2) (Liu et al., 2015). ERA5 orography is derived from RAMP2 south of 60 • S and so is closest to this benchmark but in general the slope gradient is not as steep meaning the altitude of the matched grid point is greater than the station altitude along the coastal slopes and approximately equal (within 60 m) at Neumayer on the flat Ekström Ice Shelf.
To test the sensitivity of the results to grid point collocation, the analysis was repeated with station wind data matched to each 195 of the two grid points immediately north and south of the nearest neighbour (i.e. the grid box outlined in white in the three left columns of Figure 2), with results described in Appendix A. In brief, the value of the bias for states with a primarily katabatic influence is very sensitive to this collocation test due to the sharp cutoff which occurs at the coast, but other performance characteristics and the differences between stations are robust.

Self-organising maps 200
Wind field variability local to observing stations is characterised in this study using self-organising maps (SOMs). The goal is to group periods with similar local pressure conditions, including the magnitude, gradient and orientation, without needing to create a-priori metrics to distinguish them. SOMs are a data-driven approach used to cluster multidimensional data into similar groups arranged in a 2D grid with a user-specified number of 'nodes'. A summary of the unsupervised learning algorithm applied in a climate context is given by Skific and Francis (2012). The iterative SOM approach allows non-linearities in the data SOMs in our analysis are driven using the 12-hourly ERA5 mean sea level pressure (MSLP) field over a 200×200 km grid centred upon the location of each station for 2010-2017. Each 12-hourly MSLP field in ERA5 is then matched to the closest SOM node (lowest squared difference) to determine a set of dates for each SOM. Observations and reanalysis winds are then 215 composited by these dates. 12-hourly data are used to match the approximate frequency of sonde launches at coastal stations.
SOMs are calculated using the Somoclu library for Python: https://somoclu.readthedocs.io/en/stable/download.html (Wittek et al., 2017). ms −1 though JRA-55 exhibits a higher spread of overestimated wind speeds. All three reanalyses exhibit a tendency towards negative bias with respect to ASCAT at high wind speeds (above 20 ms −1 ). This may in part relate to the localised and shortterm nature of extreme winds observed with the scatterometer. It is also possible that the representation of extremes is affected 250 by the assimilation method; for example the assimilation of extreme scatterometer winds into ECMWF analyses is sensitive to the data thinning and quality control procedures (De Chiara et al., 2017). Underestimation of high wind speeds is also a characteristic of ERA-Interim when compared with in-situ wave glider wind observations from the sub-Antarctic Southern Ocean by Schmidt et al. (2017). The mean bias across the coastal easterly domain for ERA5, MERRA-2 and JRA-55 is -0.51 ms −1 , -0.72 ms −1 and -0.06 ms −1 respectively, but sampling only wind speeds above 20 ms −1 this bias changes by -3.89 255 ms −1 , -3.88 ms −1 and -1.54 ms −1 (i.e. becomes more negative). For reference, on average across all the 10 m surface station wind speed data from Mawson, Neumayer and Dumont d'Urville (the better exposed sites), this change in the bias induced by sampling only wind speeds above 20 ms −1 would be -4.65 ms −1 , -5.13 ms −1 and -4.53 ms −1 , respectively.
Maps of the reanalysis bias with respect to ASCAT (Figure 4) indicate that the largest differences between the reanalyses and observations are found in the near-coastal region. Both ERA5 and MERRA-2 exhibit a band of negatively biased wind 260 speeds close to the coastal margins. tch: Note that regions of missing data in JRA-55 are where not enough points could be collocated (i.e. fewer than 50) between ASCAT and the reanalysis due to the reduced frequency (6-hourly) of data compared to ERA5 and MERRA-2 (hourly). Some regions of elevated bias are found close to complex or steep orography. For  It is interesting to note that although the selected domain has an overall easterly low-level wind regime, biases in the zonal and meridional components of ERA5, MERRA-2 and JRA-55 10 m winds are mixed around the coast ( Figure S5). The offshore easterly sector has much lower mean directional constancy than the onshore sector (van den Broeke and van Lipzig, 2003) meaning sb: surface wind near-surface wind directions are more variable. As a result, biases may cancel out to an extent 275 when individual wind components are calculated if the reanalysis bias is insensitive to wind direction.
Our analysis agrees with that of Carvalho (2019) who shows increasing reanalysis errors with respect to satellite wind products near the poles. Carvalho (2019) also finds that MERRA-2 generally outperforms JRA-55, CFSR, JRA-55 and ERA-Interim winds at high southern latitudes. We find that the more recent ERA5 now slightly outperforms MERRA-2 in the coastal region, consistent with the findings of Rivas and Stoffelen (2019) who report an improvement in ERA5 ocean winds relative to 280 ERA-Interim.
Some of the observed differences between ASCAT and the reanalyses may be due to inconsistencies in how scatterometer and reanalysis 10m winds are derived, although the effects of atmospheric stability are accounted for in our analysis tch: by making use of the reanalysis neutral winds rather than stability-dependent winds. One tch  Figure 13). In summary, these sources of inconsistency do have a small impact but are not able to explain large wind speed biases exceeding 2 ms −1 such as those found close to complex Antarctic orography. Coastal stations are exposed to a variety of synoptic and mesoscale flow regimes, with the largest short-term variability in winds observed offshore (Figure 1). SOM nodes for each station organise broadly into states for which the pressure gradient is more intense and ones where it is reduced ( Figure 5 and Table S1). However, the SOM nodes are not a simple continuum 305 between those two extremes; a number of nodes represent the varied orientation of synoptic pressure gradients relative to the location of the station. For example, at Dumont d'Urville (Figure 5b), nodes (1, 1) and (1, 2) have comparable large-scale pressure gradients but the low-level pressure contours associated with the offshore low are orientated along the coast for (1, 1) whereas in (1, 2) they are oriented more perpendicular to the coast. tch: This is due to the differing position of the low pressure centre with respect to the station The strongest 10 m wind speeds near each station are observed during states with the largest 310 synoptic pressure gradient ((1, 2) at Casey, (1, 1) at Mawson and Neumayer and both at DDU). The maximum in wind speed 13 is generally found onshore, which could be an indication of combined katabatic and synoptic forcing (e.g., Turner et al., 2009;Orr et al., 2014) but is also the region of maximum large-scale pressure gradient force on average (van den Broeke and van Lipzig, 2003). Each station has two or three nodes which correspond to reduced pressure gradients (Table S1), including one with high 315 pressure onshore (node (0, 1) in Figure 5). To assess the likely contribution of katabatic flow to these states, the acceleration term due to katabatic forcing K (KAT index, hereafter) is calculated for each ERA5 grid point following the approach described in van den Broeke and van Lipzig (2003). Our implementation of this approach is detailed in Appendix C.
Composites of the katabatic index ( Figure 6)  slopes favour considerable local baroclinicity. For all stations, the largest nearby composited KAT index values are for node (0, 1), i.e. during high pressure conditions. These nodes are also associated with the lowest mean total cloud cover in ERA5 (not shown), favouring long-wave cooling and a stable boundary layer, and exhibit a sharp cutoff in wind speeds at the coast. Nodes which exhibit these signs of strong katabatic influence but little synoptic influence (i.e. a low background pressure gradient) will be referenced several times in the reanalysis evaluation, so to help quickly identify them they have been marked with a 325 ' tch: [K][Kat]' in Figures 5 to 10 tch: and 7 to 10. However, these are subjectively classified and large-scale synoptic forcing in other nodes, for instance due to offshore cyclones, does not preclude the continued occurrence of katabatic winds, hence SOM states other than those marked with a ' tch: [K][Kat]' are still likely to contain a katabatic influence.

Performance at Casey station
With the lowest Pearson correlation coefficient (r) across reanalyses, sb: surface wind near-surface winds observed at Casey station 330 are poorly represented (Figure 7), especially in JRA-55 and MERRA-2 in which the topography of the Law Dome is highly smoothed (Figure 2). Large differences tch: between the observed and reanalysed wind direction are commonin both the zonal and meridional wind components between observations and reanalysis occur at Casey (Figure 7 tch: profiles). One standout is node (1, 2), a state in which the pressure contours are aligned to favour some flow around the Law Dome. r values for node (1, 2) are higher than for other nodes ( Figure D1a) largely due to better representation of wind speed variability above 15 ms −1 . In JRA-55, 335 for instance, there is considerable overestimation of sb: surface wind near-surface wind speeds below 10 ms −1 (Figure 7 Figure 7 but for DDU. tch: All shaded correlation coefficients are significant at the 99% level, according to a two-sided t-test.
tch: These three SOM nodes account for 68.5% of the total variability at DDU.

Performance at Dumont D'Urville station
At Dumont d'Urville (DDU), 10 m reanalysis winds are more closely aligned with observations than at Casey but performance characteristics vary considerably by SOM state (Figure 8 scatterplots) Figures 8m and q), which can be seen by comparing node (0, 0) and (1, 1) for the two reanalyses. tch: These reanalyses consistently tch: overestimatingoverestimate wind speeds when the pressure contours favour a more northerly flow (node (0, 0)) tch: 8g, m and 8l, r respectively) but tch: underestimatingunderestimate it when the observed low-level flow has a stronger southerly component (node (1, 1) tch: in Figures 8k and q). Root mean square error (RMSE) is also slightly higher 355 for all reanalyses in node (1, 1) compared to node (0, 0) tch: and (1, 2) ( Figure D1b). Compared to Casey, a more complex temperature profile is observed at DDU. Although not clear from the averaged temperature profiles, near-surface inversions are again common ( Figure S10) but they are also accompanied by inversions at altitudes 375 above 1000 m a.g.l., generating a distinct reverse-S shape in the temperature profile which is present in reanalyses but is smoothed and varies considerably in vertical structure between dataset. Substantial negative temperature biases occur in the lowest 1000 m in JRA-55 and MERRA-2, reaching 5 K at 100 m a.g.l. tch: (Figures 8g-l and 8m-r).

ERA5
MERRA-2 JRA-55 Figure 9. As in Figure 7 but for Mawson. tch: All shaded correlation coefficients are significant at the 99% level, according to a two-sided t-test. tch: These threee SOM nodes account for 59.3% of the total variability at Mawson.

Performance at Mawson station
Compared to other stations, performance statistics differ little between reanalyses at Mawson (Figure 9), though as before 380 ERA5 has marginally higher r in all nodes. All three underestimate the highest wind speeds, for example in node tch: (1, 1)(0, 0) ( Figures 9a, d, g). This underestimation of elevated wind speeds at Mawson is also a feature of the Met Office Unified Model (UM) which was not rectified by switching to a very high horizontal or vertical grid resolution (Orr et al., 2014).
As at DDU, tch: katabatic statesthe katabatic state shown tch: atfor Mawson ( tch: (0, 1) and (1, 0)) tch: areis characterised by a considerably reduced IQR of reanalysis wind speeds with respect to observations ( Figure D1c); when an intense pressure gradient is absent, 385 reanalyses appear to fall into a regime of underestimating the large local wind speed variability. One interpretation is that there are two distinct flow regimes at Mawson. If the large-scale pressure gradient is strong enough, the onshore and offshore wind field become continuous, whereas when it is weaker terrestrial (likely katabatic) winds are cut off abruptly at the coast and the representation of the winds becomes sensitive to the placement and turbulent characteristics of the katabatic jump.
The easterly jet at Mawson is similar to DDU (Figure 9) tch: , peaking with high synoptic pressure gradients in states (0, 2) and (1, 1) and 390 accompanied by a shallow southerly layer. As with DDU, correlation coefficients against vertical wind measurements vary by SOM node at Mawson. Interestingly, high and consistent r values are found throughout the 2500 m profile for each of the three reanalyses in node tch: (0, 2)(0, 0) during which the pressure gradient is high but the southerly layer is shallow as a result of the orientation of the low pressure centre offshore, directing geostrophic flow of maritime origin towards Mawson station. tch: By contrast, when a deeper and stronger southerly layer prevails in node (1, 1), the r values are reduced. This may relate to the occurrence of temperature 395 inversions which are poorly resolved by the reanalyses ( Figure S10). tch: All three reanalyses underestimate the strength of the u-wind in strong jet cases (1, 1), although the structure of the v-wind and height of the core of the low-level jet is more realistically represented in JRA-55.
The temperature profile at Mawson is comparable to DDU in its reverse S shape. Again, this is a feature that varies greatly between reanalyses; the structure is most realistic in JRA-55 and ERA5. This may explain why the jet structure (with an elevated easterly core and lower southerly core) is also more realistic in those reanalyses compared to MERRA-2 for which 400 this structure in the vertical wind profile is not seen. As with other stations, JRA-55 and MERRA-2 exhibit a cold bias in the lowest 1000 m, with that of JRA-55 equalling or exceeding 5 K at the surface.

ERA5
MERRA-2 JRA-55 Figure 10. As in Figure 7 but for Neumayer. tch: All shaded correlation coefficients are significant at the 99% level, according to a two-sided t-test. tch: These three SOM nodes account for 61.3% of the total variability at Neumayer.
23 3.2.5 Performance at Neumayer station As described in Section 2.3.1, Neumayer has quite different local terrain characteristics compared to other stations, being situated on a flat ice shelf with comparably small orographic height differences between reanalyses and reality. sb: surface windNear-405 surface wind disagreements are much smaller at Neumayer than at the other stations considered here ( Figure 10). As a result, the lowest mean RMSE values are found here ( Figure D1d). ERA5 appears to be the best performing reanalysis by a larger margin than at Mawson. By contrast, lower r values and higher RMSE are found for all SOM nodes in JRA-55 ( Figure D1d).
There is little discernible difference in the scatterplot characteristics between SOM nodes at Neumayer.
Consistent with better reanalysis performance against surface station winds, both reanalysis wind profiles and reanalysis low-level flow passing over the complex coastal terrain east of Neumayer. The jet structure is well represented, including a shallow southerly flow. However, the core of the easterly jet, seen most clearly in tch: nodesnode (1, 1) tch: and (1, 2) tch: in Figure   415 10, is consistently underestimated by all three datasets. Baroclinicity and the coastal slopes nearby are thought to introduce ageostrophic motions at Neumayer (Kottmeier, 1986;Klöwer et al., 2014) though the katabatic pressure gradient force is not as large over the ice shelf as further inland (van den Broeke and van Lipzig, 2003). Gorodetskaya et al. (2020) show that a slight negative ERA5 wind speed bias is also a characteristic of low-level jets occurring in atmospheric river conditions at Neumayer; our analysis indicates that this bias applies more generally to strong wind conditions, for example as in node (1, 1).

Discussion
In this paper, we have evaluated the representation of Antarctic coastal winds in three current reanalyses. The analysis aims to quantify representation of the mean state and short-term variability of coastal winds, making use of the few observational datasets available within the coastal easterly domain. There is a great deal in common between reanalyses in the representation of time-mean coastal winds and their variability. These similarities and the differences which do exist point to some critical 425 processes which are important for representation of the coastal tch: easterlieswinds.
Firstly, the representation of orography is especially important at the coast and has a large impact on reanalysis performance.
As shown in Section 2.3.1, the slope of the coastal orography in reanalyses is smoothed, leading to large differences in surface height at the latitudes of most coastal stations, with the exception of Neumayer where the local terrain is flat. At Casey, the reanalysis performance is poor across the board, especially for the coarser MERRA-2 and JRA-55, suggesting the local 430 orography (in particular the Law Dome) has a large effect on the winds, including those well above the surface layer. Whereas reanalysis winds are quite realistic at Neumayer, correlation coefficients only improve at Casey when the large-scale flow is relatively unimpeded by the Law Dome. The comparison with ASCAT also points to a critical role for orography in the coastal margins; reanalysis performance with respect to ASCAT declines with increasing proximity to the coast and biases are especially large close to complex orography such as the Transantarctic Mountains, Law Dome and around Enderby Land.

435
Orographic processes such as barrier and tip jet formation are likely to be important in the Antarctic coastal margins and sources of marine wind speed biases in reanalyses.
A second and related theme is the difficulty in characterising low-wind states dominated by katabatic forcing when a strong large-scale synoptic pressure gradient is absent (nodes marked with a ' tch: [K][Kat]' in Figures 7, 8 and 9). This is clearest at DDU and Mawson which are subject to semi-permanent katabatic winds. Vignon et al. (2019) show how a typical feature of 440 ERA-Interim and ERA5 is an underestimated range between the 5 th and 95 th percentiles of wind speed in the lowest 200-300 m. Our analysis indicates this effect may be greatly enhanced during these tch: [K] [Kat] states; at DDU this mostly appears as an overestimate of low wind speeds whereas at Mawson high wind speeds are underestimated. The exact source of this misrepresentation is not clear from this analysis but the implication is that winds in the coastal margins are more variable than reanalyses suggest. A caveat of this analysis is that katabatic winds are a highly terrain-dependent and locally variable 445 phenomenon at the coast; the sign of the bias is quite sensitive to the collocation. Furthermore, the contribution of tch: momentum from katabatic flow to maritime winds responsible for coastal currents remains unquantified, but our analysis indicates that katabatic regimes may be a substantial source of model error.
A third point of interest is that temperature profiles at coastal stations differ between reanalyses, and although this research focuses on winds these differences could be important for tch: the coastal easterliesthe baroclinic conditions which influence coastal low-wind katabatic states, may be in part related to these temperature biases and in part to the representation of temperature inversions. A dual inversion structure as shown for Mawson and DDU in Figure S10 is common for the Antarctic coastline (Truong et al., 2020) and poses a challenge for faithful representation of near-coastal thermodynamic profiles but its role in the representation of wind profiles warrants further investigation.

Implications for use of reanalysis datasets 460
Reanalyses are capable of representing important features of the coastal tch: easterlieswind field and tch: theirits variability, including the existence of a jet feature peaking during strong synoptic forcing. Many of the differences against observations are consistent between datasets, including underestimated sb: surface wind near-surface winds above 20 ms −1 , underestimated interquartile range in states with reduced pressure gradients where katabatic winds dominate, poor representation at Casey where a local orographic feature plays a major role and improved performance at Neumayer over relatively flat surface conditions. 465 tch: Alongside the detailed breakdown of performance statistics across various metrics, rankings for each reanalysis by station and by metric (using only surface wind speed) are given in Appendix D. These indicate that ERA5 exhibits the best performance overall, especially for r (as with ASCAT). Higher resolution in ERA5 also supports more realistic coastal orography; steeper coastal slopes are in turn likely to be important for realistic terrain-driven baroclinicity and katabatic forcing (Fulton et al., biases in the lowest 1000 m a.g.l. and a decline in wind speed r values near the top of the easterly jet region in some SOM states. However, JRA-55 and MERRA-2 exhibit lower mean bias and bias in interquartile range. Negative bias is especially large at high wind speeds in ERA5 when compared with both station measurements and ASCAT. The contribution of this substantial and time-varying bias to long-term trends and wind stress in the coastal sector warrants further investigation but these results suggest that using ERA5 to analyse coastal extremes requires careful consideration of the effect that a systematic 475 negative bias could have upon the results.
tch: There are several reasons why ERA5 may be most useful as a reference dataset in the coastal easterly sector. Firstly, surface r values against ASCAT and station measurements are on average highest for ERA5, indicating the most faithful representation of variability in the 10 m winds on short timescales.
Secondly, higher resolution in ERA5 supports more realistic coastal orography which is critical in coastal regions; JRA-55 and MERRA-2 coastal orography is not as steep and theory suggests that the slope is likely to be important for both katabatic winds and for a deeper baroclinic layer in the coastal zone. Lastly,

480
the temperature profile appears to be most realistic in ERA5, with JRA-55 and MERRA-2 exhibiting substantial cold biases in the lowest 1000 m a.g.l. and a decline in wind speed r values near the top of the easterly jet region in some SOM states. An important caveat in using ERA5 is that it still exhibits the large negative time-mean wind speed biases, shown for both ASCAT and station measurements and especially pronounced at high wind speeds. The contribution of this substantial and time-varying bias to long-term trends and wind stress in the coastal sector warrants further investigation but these results suggest that using ERA5 to analyse coastal extremes requires careful consideration of the effect that a systematic negative bias could have upon the results.

485
ASCAT observations and sonde data were assimilated into all three reanalyses, meaning only the 10 m wind observations are quasi-independent in our evaluation. The assimilated observations were likely important for reanalysis performance and so further evaluation with unassimilated observations would be valuable. For example, the effect of assimilating new Antarctic dropsonde observations from the Concordiasi field experiment on ECMWF forecasts was found to be larger away from the coastal regions where most of the existing radiosondes are launched (Boullot et al., 2016) and additional sonde observations 490 from the Year of Polar Prediction improved wind speed forecasts over West Antarctica (Bromwich et al., 2022).

Conclusions
The representation of tch: winds within the region of the Antarctic coastal easterlies in the ERA5, MERRA-2 and JRA-55 reanalyses has been evaluated by comparison with in-situ and satellite observations, including summertime ASCAT observations over the marine sector and observations from four coastal stations. For each station, the sensitivity of reanalysis performance 495 to changes in the driving flow regime was analysed using SOMs derived from ERA5 MSLP fields.
A comparison with ASCAT scatterometer winds shows that wind speed variability within the region of the coastal easterlies is well represented by the reanalyses, with correlation coefficients of 0.91 in ERA5, 0.89 in MERRA-2 and 0.85 in JRA-55.
Generally, sb: surface wind near-surface wind speeds offshore are biased slightly low across the whole coastal easterly sector in ERA5 and MERRA-2 whereas JRA-55 biases are closer to zero on average but have large local variations. Correlation 500 coefficients decrease with proximity to the coast and the ASCAT fields reveal near-coastal wind features (especially close to complex orography) whose magnitude is not well captured by the reanalyses. All three reanalyses exhibit larger negative biases relative to scatterometer data at wind speeds exceeding 20 ms −1 .
Reanalysis performance against observations is much more sensitive to different flow regimes (characterised with SOMs) at coastal slope stations (in particular Mawson and DDU) than over the flatter terrain of Neumayer. Time series correlation 505 coefficients (r) between both surface and upper air wind speeds and collocated reanalysis wind speeds are consistently higher in states with strong synoptic forcing and especially when low-level flow is directed onshore. By contrast, r values consistently drop for states with limited synoptic forcing but favourable conditions for katabatic forcing. This is in part due to the lower wind speeds but also because the reanalyses' interquartile range is lower than observations at both Mawson and DDU during those states. Performance against observations from Casey is poor due to the effect of the Law Dome on the low-level flow.

510
Representation of orography is likely critical to variations in reanalysis fidelity by state and station. MERRA-2 and JRA-55 have large cold biases in the lowest 1500 m which may be important for the realism of low-level coastal tch: easterlieswinds.
Overall, ERA5 has the highest sb: surface wind near-surface wind r values compared to both station measurements and ASCAT over the coastal easterly sector. High r values are generally mirrored by low RMSE. A more realistic temperature profile, better characterisation of variability and more faithful representation of orography gives greater confidence in ERA5 as a benchmark 515 dataset for tch: the coastal easterliesAntarctic coastal winds. It should be emphasised, however, that performance is similar between reanalyses and many key deficiencies are shared, including relatively large biases against ASCAT near the coast and at high wind speeds, poor performance close to the Law Dome and underestimation of the wind speed variability when synoptic forcing is reduced and katabatic forcing is important.
These results shed light on some of the challenges associated with evaluating model-based datasets in the Antarctic coastal 520 region and underscore the role of orography. Furthermore, they demonstrate the usefulness of an evaluation approach designed to interrogate reanalysis or model performance under varied atmospheric states, given the high variability of the coastal sector.
This SOM-based methodology offers an approach to reanalysis evaluation that is process-oriented and could be applied to other reconstruction datasets or assessment of climate models. In future work the insights from this analysis will be used to For the comparison with station measurements, all results were recalculated after shifting the collocated reanalysis gridpoint one point to the north and south.
Although this shift causes a slight degradation in correlation coefficients for all reanalyses, stations and SOM nodes, the main results reported in Sections 3.2.2 to 3.2.5 are robust, except for a large impact upon the value of the bias for wind speeds below 10 ms −1 in states and at stations where the synoptic forcing is reduced but katabatic forcing is favoured. Low-wind 540 states in steep-slope coastal regions therefore appear to be highly sensitive to collocation. For example, in ERA5 for SOM node (0, 2) at DDU the bias is 0.77 ms −1 , but shifting the collocated point south increases this to a 5.19 ms −1 bias (with a similar effect on low winds in other states) and, conversely, shifting it north gives a -0.9 ms −1 bias. Large wind speed gradients occur at the Antarctic coast, especially during katabatic-dominated states, meaning winds upslope (south) of surface stations may be much stronger than those just offshore (north). Estimates of reanalysis or model bias in the representation of katabatic 545 winds based upon point collocation with coastal station measurements should therefore be treated with caution. Our focus is primarily upon differences between nodes, stations and reanalyses.

Appendix B: Selection of SOM configuration
A variety of SOM configurations were tested ( Figure B1). Node correlation and pair correlation were calculated based on the approach of Gibson et al. (2017). Node correlation is defined as the mean correlation between the data points comprising 550 individual SOM nodes and the data points of timesteps to which they were assigned (i.e. a reanalysis timestep is assigned to a SOM node based on the mean squared difference and then the correlation is calculated between all data points in that timestep and all data points in the SOM node). The pair correlation is defined as the mean correlation between all possible pairs of SOM nodes. The node correlation is therefore intended as a measure of how realistic SOM nodes are (larger values are desirable) whereas pair correlation is a measure of how similar they are to each other (lower values are desirable). It should be 555 noted that pair correlations near one for configurations with few nodes are a result of the background large-scale conditions of the Antarctic; high pressure over the continent and low pressure offshore within the circumpolar trough, with this dipole not differing greatly between nodes in Figure 5 despite varying patterns. Only at Neumayer is there some evidence in Figure 5 of this dipole reversing.
SOMs for Casey, Mawson and DDU were found to be relatively insensitive to the number of nodes, with very little change 560 in the above statistics until around a 3 × 6 configuration ( Figure B1). Neumayer, on the other hand, is much more sensitive to the number of SOM nodes. This is likely because it is situated within a region of high MSLP variability (near the northern edge of the peak 12-hourly wind speed standard deviation in Figure 1) compared to other stations where the local flow is much more constrained by orography. It could be argued then that a 3 × 5 configuration would be preferable for Neumayer. However, the performance of the reanalyses was found not to vary by node at Neumayer even for SOM grids with many nodes. For ease 565 of interpretation, therefore, we opt for a grid with fewer nodes. Inclusion of additional SOM nodes at other stations was not Figure B1. Indicators of SOM realism and node uniqueness for different configurations (grid sizes). Solid lines are the mean node correlation and dashed lines the mean pair correlation for Casey (green), DDU (red), Mawson (blue) and Neumayer (orange).
found to add much value to the analysis except in distinguishing different varieties of extreme wind events, which is beyond the scope of the current work but may be useful for a more targeted analysis in future.
tch: The effect of the length of the time series upon the selection of SOM configuration was also tested. Repeating the analysis used to produce Figure B1 yielded very similar results with the time series artificially reduced from 8 years to 1. MSLP 570 composites were also found to be very visually similar to Figure 5, including when driven with hourly data instead of 12-hourly (not shown).
Overall, A SOM configuration of 2 × 3 was deemed to be capable of capturing critical differences in reanalysis performance (associated with katabatic and synoptic states as well as variations in flow orientation) without being too large for the reader to interpret easily.

575
Appendix D: Detailed SOM state statistics As shown in the paper for Casey station (Section 3.2.2), a visual comparison of scatterplots between SOM nodes is arguably more informative than summary statistics given the regime behaviour in the winds. To supplement this analysis, however, we here provide a ranking of reanalyses based upon four performance metrics calculated from station near-surface wind measurements, the correlation coefficient (r), root-mean square error (RMSE) mean bias (MB) and bias in the interquartile range 590 (IQRB).
We rank the performance of each reanalysis for each station by calculating each statistic separately for each SOM node ( Figure D1). Each reanalysis is assigned a ranking for that node and metric (1 for the best, 3 for the worst), which is then weighted by the frequency of that node. Summing the rankings for each metric gives a final score for that station. By this measure, the overall rankings for each station are: 595 data are produced by Remote Sensing Systems and sponsored by the NASA Ocean Vector Winds Science Team. Data are available at http://www.remss.com; thanks to ICDC, CEN, University of Hamburg for data support. tch: We also thank two anonymous reviewers whose insightful comments helped to improve the paper.