1. Introduction
Surface atmospheric variables such as temperature, pressure, and wind speed in the Antarctic are notoriously difficult to simulate, yet they remain some of the most pertinent weather phenomena with regards to personnel safety and operational logistics (i.e., shipping and aviation) in the region. Strong temperature inversions develop over the Antarctic continental surface when the absorption of solar radiation is low (due to high albedo) and there is a disparity between snow-surface and atmospheric emissivity (Hudson and Brandt 2005). These temperature inversions can lead to stratification of the atmosphere whereby air aloft is warmer than air below, causing the near-surface air to become negatively buoyant, which, when coupled with steep, sloping terrain, can lead to the generation of fast gravity-driven katabatic winds (Connolley 1996).
Complicating forecasts of the surface radiative balance are the varying properties of land surfaces in the region. Ice and snow covered surfaces have different albedos, as well as different radiative or thermal properties depending on the level of melt (Grenfell et al. 1994; Kuipers Munneke et al. 2011). Sea ice presents significant challenges to forecasts due to its moderating influence on exchanges of heat and momentum between ocean and atmosphere (Maykut 1978), the sea ice–albedo feedback mechanism (Hall 2004) and thermal insulating properties (Maykut and Untersteiner 1971), and the subdaily time scales at which sea ice can freeze, melt and move (changing the albedo and radiative properties of the model grid cell). This can lead to forecasting biases in regions of large-scale sea ice growth and decay, including polynyas and at the sea ice edge. Furthermore, both simulated and observed cloud cover has a substantial effect on the modeled surface radiative balance, where overcast skies may increase downward longwave flux by as much as 80 W m−2, irrespective of season (Yamanouchi and Kawaguchi 1984). As such, positive (negative) biases in cloud cover will likely increase (decrease) incident longwave radiation at the surface, causing a warming (cooling) effect.
The complexity of the Antarctic topography and its representation in models presents challenges in predicting wind regimes, whereby steep, sloping terrain and cold boundary layer outflow aids in the production of gravity-driven katabatic winds (Connolley 1996). Furthermore, these complexities along the continental margin characterize the regions of confluence responsible for persistent katabatic winds (Parish and Bromwich 1987). Excessive winds can lead to enhanced mechanical mixing, resulting in a warming effect and underrepresentation of static stability at the surface (Nigro et al. 2017). Underrepresentation of coastal topographical complexity due to poor model resolution therefore has numerous downstream impacts on the predictability of forecast parameters such as surface winds and temperatures.
The Year of Polar Prediction (YOPP; see Jung et al. 2016) special observing periods (SOPs) are a concerted international effort to increase the number of observations made in the polar regions during intensive and coordinated campaigns. These additional observations are made available to all national weather services via the WMO Global Telecommunications System (GTS) for the purposes of data assimilation and model verification by any center using the system. Model evaluation is then conducted in order to understand the potential increase in predictability of the polar regions based on the increase in observational data. The austral summer SOP occurred between 16 November 2018 and 15 February 2019 (to span the period of greatest operational activity in the Antarctic) and effectively doubled the number of radiosonde releases as well as increased drifting buoy deployments (Bromwich et al. 2020); the data from which was found to yield the greatest forecast improvement for deep cyclones near the Antarctic coast. However, Bromwich et al. (2020) also found that Southern Hemisphere forecast skill still lagged behind that of the Northern Hemisphere. This study performs in intercomparison between operation NWP models during the Antarctic summer SOP, following on from the work of Bromwich et al. (2020) and comparable studies in the Northern Hemisphere (Køltzow et al. 2019).
2. Data and methods
Height difference between each model and ERA5 (regridded onto each model native grid). These differences are used to height-correct 2-m temperature and surface pressure. Note: ARPEGE-SH and ARPEGE-SH-GELATO share the same topographical field.
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
Height difference between each model and ERA5 (regridded onto each model native grid). These differences are used to height-correct 2-m temperature and surface pressure. Note: ARPEGE-SH and ARPEGE-SH-GELATO share the same topographical field.
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
Height difference between each model and ERA5 (regridded onto each model native grid). These differences are used to height-correct 2-m temperature and surface pressure. Note: ARPEGE-SH and ARPEGE-SH-GELATO share the same topographical field.
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
Configuration details for the models evaluated in this study. Model versions are as was operational during the study period.
The global version of the Australian Community Climate and Earth System Simulator (ACCESS-G; Puri et al. 2013) is used operationally by the Australian Bureau of Meteorology to provide forecast guidance in support of the Australian Antarctic Division (AAD). ACCESS is an atmosphere-only model with a binary “slab” sea ice prescribed at the beginning of the forecast and unevolved throughout. ACCESS-G forecast quality is sensitive to the few observations it assimilates in the Antarctic (Soldatenko et al. 2018) and has been shown to produce slower modeled winds at the surface associated with a warm surface temperature bias at the coastal margins, with subsequent effects in the surface pressure and geopotential height fields associated with these biases, particularly over the ocean (Schroeter et al. 2019). ACCESS-G features no documented polar enhancements known to the authors and the assessment of the model’s performance over the Antarctic has been limited to date (see Schroeter et al. 2019). At the time of this study the next version of ACCESS-G, APS3, had come into operation; however, the previous version, APS2, was the version used during the SOP, hence its inclusion in this study.
The European Centre for Medium-Range Weather Forecast’s Integrated Forecast System (ECMWF-IFS) is developed by ECMWF and was configured for the YOPP special observing periods (Bauer et al. 2020). The open access, YOPP-specific dataset is generated from the unperturbed, deterministic control forecast of the 51-member operational coupled ensemble and provides 3-hourly (6-hourly) resolution out to day 7 (day 15). A parallel research suite identical in configuration to the operational suite was run with additional diagnostic outputs at higher temporal frequency for the first 48 h to allow users to break down the change of state into the contribution of each physical process from the physical and dynamic atmospheric tendencies (see Bauer et al. 2020 for further detail). While the ECMWF-IFS system is used to produce ERA5; however, ECMWF-IFS is a forecast model and thus does not have access to the additional observations and postprocessing conducted in producing the ERA5 reanalysis. There has been limited verification of the model over the Antarctic to date; however, Arctic evaluation has found the model to have a cold daytime surface bias, slow 10-m winds and a warm atmospheric temperature bias at 700 hPa (Bauer et al. 2020; Køltzow et al. 2019).
The Global Deterministic Prediction System (GDPS: Environment and Climate Change Canada 2018) is a fully coupled atmosphere–ocean–sea ice forecast model, the data of which was made available to the YOPP community. Initial evaluation of GDPS over the Antarctic for the summer SOP period found a systematic cold bias of surface air temperature, in part due to an overprediction (underprediction) of clear sky (cloudy) conditions, and underestimation (overestimation) of strong winds (weak winds, associated with night inversions) (Bromwich et al. 2020).
PolarWRF as implemented by the Antarctic Mesoscale Prediction System (AMPS) is used by the U.S. Antarctic Program as well as numerous other national Antarctic programs (Powers et al. 2012). PolarWRF features a number of polar enhancements, including a fractional sea ice component in which the surface layer scheme is called twice; once for completely frozen conditions and once for completely open water to determine fluxes and other terms given by a weighting of the results by the sea ice fraction. Other enhancements include adjustments to the thermal and radiative properties of sea ice- and snow-covered surfaces, and other improvements to the planetary boundary layer (PBL) scheme. AMPS has been shown to have positive wind speed biases at the coast which were attributable to the meridional component in prior versions (pre-PolarWRF) and, in part, to the topographic smoothing of the model (Bromwich et al. 2005). PolarWRF also has been shown to have a cold surface temperature bias near sea level, which may be the result of radiative imbalance (Nigro et al. 2017).
The Action de Recherche Petite Echelle Grande Echelle–Southern Hemisphere (ARPEGE-SH) is a model configuration developed by the National Centre for Meteorological Research (Météo-France). ARPEGE-SH is based on the ARPEGE global version (Pailleux et al. 2015), with a high-resolution grid structure located over the Antarctic. ARPEGE-SH is uncoupled, with climatologically based sea ice temperatures remaining constant during the forecast (Bazile et al. 2020). Initial verification of ARPEGE-SH against its global parent model, ARPEGE, has found a positive impact on atmospheric temperature RMSE which is best observed beyond 24 and 60 h against radiosondes and ERA5, respectively (Bromwich et al. 2020). Bromwich et al. (2020) conclude that this improvement can be attributed to the increased horizontal resolution of the model and perhaps an increase in the number of assimilated radiosonde temperature observations in the boundary layer. ARPEGE-SH-GELATO is identical to ARPEGE-SH in all respects except coupling to the GELATO-1D sea ice model, which resolves a single category of sea ice without horizontal processes, 10 equal-thickness ice categories (defined by their enthalpy), and a single layer of possible sea ice snow cover (see Bazile et al. 2020 for further information). Although verification of the ARPEGE-SH-GELATO system is limited to date, initial experiments have found an improvement in temperature RMSE over the Antarctic below 700 hPa compared to the ECMWF operational analysis (Bazile et al. 2020).
Mean error (ME) for the study period plotted as biases from the reference analysis (ERA5). Models are indicated with different colors, cell classifications with marker style, and forecast lead with opacity. Rightmost models (ARPEGE-SH-GELATO, ECMWF-IFS, and GDPS) are coupled.
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
Mean error (ME) for the study period plotted as biases from the reference analysis (ERA5). Models are indicated with different colors, cell classifications with marker style, and forecast lead with opacity. Rightmost models (ARPEGE-SH-GELATO, ECMWF-IFS, and GDPS) are coupled.
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
Mean error (ME) for the study period plotted as biases from the reference analysis (ERA5). Models are indicated with different colors, cell classifications with marker style, and forecast lead with opacity. Rightmost models (ARPEGE-SH-GELATO, ECMWF-IFS, and GDPS) are coupled.
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
3. Results
The results that follow are communicated in terms of positive or negative biases [i.e., where a model forecasts a value higher or lower than the verifying analysis (ERA5), respectively]. It is important to note that while the term “error” may be used to describe a performance metric, ERA5 is an imperfect representation of reality. Hence, differences between a model and ERA5 should rather be interpreted as a “difference” or “bias” with regards to ERA5.
a. Mean error and standard deviation of error
Figure 2 and 3 show the relationship between mean error (ME) and standard deviation of error (SDE) for each model and study parameter; this approach is similar to that of (Køltzow et al. 2019) and provides insight into the variability of model performance about the mean bias, which may be subject to issues of centrality in polarized positive/negative bias fields. The results that follow are communicated in terms of bias at a given forecast length and bias growth as a measure of how that bias developed from analysis out to longer forecast lengths; similarly, results are discussed in terms of cell classifications over a given surface type.
All models are generally in good agreement (±1 K) with the reference analysis over open ocean for 2-m temperature, with low ME (Fig. 2a) and SDE (Fig. 3a); although PolarWRF exhibits a stronger negative bias than the other models. In this result it appears that the lower resolution of ACCESS-G yields a lower mean error than expected; however, when viewed in concert with SDE (Fig. 3a) there is comparable variability in the range of errors about 0, suggesting a dipole effect which is cancelling out errors. All models (except ECMWF-IFS) are cool-biased over both low and high concentration sea ice, the coolest of which is ARPEGE-SH up to −3 K ME and between 2 and 2.5 K SDE for high concentration sea ice. Most models are cool biased at the coast, with PolarWRF and ARPEGE-SH approaching −2 and −3 K at longer forecast lengths, respectively. All models (except ECMWF-IFS) are warm-biased over land at analysis, with ACCESS-G and ARPEGE-SH-GELATO approaching +2 and +3 K around 36–48 h. SDE increases in all models and over all cell classifications with forecast length (Fig. 3a).
Standard deviation of error (SDE) for the study period plotted as biases from the reference analysis (ERA5). Models are indicated with different colors, cell classifications with marker style, and forecast lead with opacity. Rightmost models (ARPEGE-SH-GELATO, ECMWF-IFS, and GDPS) are coupled.
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
Standard deviation of error (SDE) for the study period plotted as biases from the reference analysis (ERA5). Models are indicated with different colors, cell classifications with marker style, and forecast lead with opacity. Rightmost models (ARPEGE-SH-GELATO, ECMWF-IFS, and GDPS) are coupled.
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
Standard deviation of error (SDE) for the study period plotted as biases from the reference analysis (ERA5). Models are indicated with different colors, cell classifications with marker style, and forecast lead with opacity. Rightmost models (ARPEGE-SH-GELATO, ECMWF-IFS, and GDPS) are coupled.
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
Considering surface pressure, we see in Fig. 2b that the models are generally in good agreement (±0.75 hPa) over open water and both classifications of sea ice and Fig. 3b shows that the SDE of surface pressure over these cell classifications follow a similar trajectory out to longer forecast lengths for all models. Of note is the comparative performance over high-concentration sea ice between ARPEGE-SH and ARPEGE-SH-GELATO. These models are identical in all aspects except sea ice coupling, which is shown to have a positive impact on surface pressure over high concentration sea ice in ARPEGE-SH-GELATO, tending toward a ME of 0 hPa as the forecast length increases rather than continuing to increase as it does in ARPEGE-SH.
Coastal surface pressure ME is well constrained to ±1 hPa throughout the model cohort; however, positive error growth in GDPS is greater than in other models (Fig. 2b). SDE increases with forecast length in all models, following a similar trajectory to non-land cells (Fig. 3b). For all models except ACCESS-G, ME over land is negatively biased and exhibits higher (but stable) SDE than other cell classifications for each model. Model surface pressure ME/SDE behaviors vary between models over land. For example, positive pressure biases in ACCESS-G weaken (improve) by around 0.4 hPa from analysis to 12 h, before becoming negatively biased by the same magnitude at 48 h. ARPEGE-SH-GELATO has the lowest negative continental ME at analysis; however, this increases by nearly 0.5 hPa by 48 h with SDE of around 9 hPa, the highest of any model examined in this study (Fig. 3b).
All models analyzed here typically have low ME wind speed scores over all cell classifications (±1.5 m s−1, Fig. 2c), with the exception of ACCESS-G, which has persistently slower modeled winds over non-land cells, and ARPEGE-SH-GELATO, which has an increasing negative bias over land. Figure 3c shows that while ME may be low for the majority of cell classification, SDE increases proportional to forecast length in all cases (Fig. 3c), with coastal wind SDE increasing faster in the highest-resolution models, ARPEGE-SH and ARPEGE-SH-GELATO.
The 500-hPa geopotential height SDE increases with forecast length for all models and parameters (Fig. 3d). Geopotential height forecasts over the ocean are well constrained to within ±5 m from the reference analysis, with only as subtle upward trend (increasing positive ME) in all models except ACCESS-G (Fig. 2d). Forecasts over both sea ice classifications are positively biased in all models (i.e., the 500-hPa surface is elevated above that of the reference analysis), with a tendency for higher biases over sea ice with high sea ice concentration than that of low concentration at each forecast step. Despite several models starting at analysis with low or negative coastal ME bias in 500-hPa geopotential height, all models show positive and growing ME biases as the forecast progresses. Biases in geopotential height over land vary between models over time, with vacillations in negative biases (ACCESS-G, PolarWRF, and ARPEGE-SH-GELATO), stable positive biases or substantial positive error growth (+10 m, GDPS).
b. 2-m temperature
Figure 4 shows the mean bias in 2-m temperature at 0 h (analysis, first column) and 48 h (second column), and the correlation between 0- and 48-h biases (third column). All models except ECMWF are positively (warm) biased over land at analysis and remain so out to 48 h; approaching positive biases in excess of 2.5 K which are significant with respect to the ERA5 anomaly over the study period. There is weak-to-moderate correlation between biases at analysis and those at 48 h over land, with the exception of ARPEGE-SH and ARPEGE-SH-GELATO which show slightly stronger correlation over the Ross Ice Shelf (Figs. 4i,l). Biases are predominantly negative on or near the coast in all models, extending northward over the ocean in GDPS, ARPEGE-SH-GELATO, ARPEGE-SH, and PolarWRF in increasing extent. ACCESS-G shows low error over the ocean due to a strong seasonality of performance as well as smoothing of errors at lower resolution. Although errors in ACCESS-G are low over the ocean, they are moderate-to-strongly correlated between analysis and 48-h forecast (Fig. 4c), likely due to the prescribed sea ice field in the model. ARPEGE-SH-GELATO has lower magnitude ME than ARPEGE-SH over both land and ocean, in addition to a substantially decorrelation between analysis compared to its uncoupled counterpart (Figs. 4g–l).
2-m temperature mean bias at (left) 0 h (analysis) and (center) 48 h, and (right) the correlation coefficient between 0- and 48-h biases. Stippling (if illustrated) indicates significance from ERA5 anomalies across the study period (α = 0.05).
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
2-m temperature mean bias at (left) 0 h (analysis) and (center) 48 h, and (right) the correlation coefficient between 0- and 48-h biases. Stippling (if illustrated) indicates significance from ERA5 anomalies across the study period (α = 0.05).
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
2-m temperature mean bias at (left) 0 h (analysis) and (center) 48 h, and (right) the correlation coefficient between 0- and 48-h biases. Stippling (if illustrated) indicates significance from ERA5 anomalies across the study period (α = 0.05).
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
c. Surface pressure
Figure 5 shows the mean bias in surface pressure at 0 h (analysis, first column) and 48 h (second column), and the correlation between these biases (third column) for each model. Biases are higher over land than they are over the ocean (irrespective of sign) at analysis for most models, with the lower resolution models (ACCESS-G and GDPS) largely insulated from topographical differences experienced at higher resolutions, which contribute to synoptic-scale bias profiles (Figs. 5a,p); additionally, both of these models have positive surface pressure biases over the Ross Sea and significant negative biases over portions of East Antarctica (120°E). It is noted that there are polar artifacts impacting performance in ACCESS-G, which upon deeper examination of the raw data were found to be attributed to a persistent, erroneous surface pressure bias close to the pole (Figs. 5a–c) that was not previously observable at a lower resolution version of this analysis (not shown). ARPEGE-SH and ARPEGE-SH-GELATO have strikingly different surface pressure bias profiles, despite differing only in coupling. However, ARPEGE-SH-GELATO having continental biases of higher magnitude than ARPEGE-SH, these errors are not correlated as they are in ARPEGE-SH (Figs. 5g–l).
Surface pressure mean bias at (left) 0 h (analysis) and (center) 48 h, and (right) the correlation coefficient between 0- and 48-h biases. Stippling (if illustrated) indicates significance from ERA5 anomalies across the study period (α = 0.05).
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
Surface pressure mean bias at (left) 0 h (analysis) and (center) 48 h, and (right) the correlation coefficient between 0- and 48-h biases. Stippling (if illustrated) indicates significance from ERA5 anomalies across the study period (α = 0.05).
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
Surface pressure mean bias at (left) 0 h (analysis) and (center) 48 h, and (right) the correlation coefficient between 0- and 48-h biases. Stippling (if illustrated) indicates significance from ERA5 anomalies across the study period (α = 0.05).
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
d. 10-m wind speed
Figure 6 shows the mean bias in 10-m wind speed at 0 h (analysis, first column) and 48 h (second column), and the correlation between these biases (third column). ACCESS-G is significantly negatively (slow) biased over the ocean and up to the coastal margin at both analysis and at 48 h (Figs. 6a,b); these biases do not grow substantially, but are strongly correlated around the sea ice edge. This behavior is not present in the other models, which have more sophisticated sea ice componentry (be it coupling or otherwise) than prescribed as it is in ACCESS-G. PolarWRF, ARPEGE-SH, ARPEGE-SH-GELATO, and GDPS are negatively (slow) biased over land at analysis (Figs. 6d,g,p), all three models maintaining persistent slow continental winds at 48 h (Figs. 6h,k,q). Of the models examined, PolarWRF develops positive (fast) wind speed biases at the coastal margins.
Wind speed mean bias at (left) 0 h (analysis) and (center) 48 h, and (right) the correlation coefficient between 0- and 48-h biases. Stippling (if illustrated) indicates significance from ERA5 anomalies across the study period (α = 0.05).
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
Wind speed mean bias at (left) 0 h (analysis) and (center) 48 h, and (right) the correlation coefficient between 0- and 48-h biases. Stippling (if illustrated) indicates significance from ERA5 anomalies across the study period (α = 0.05).
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
Wind speed mean bias at (left) 0 h (analysis) and (center) 48 h, and (right) the correlation coefficient between 0- and 48-h biases. Stippling (if illustrated) indicates significance from ERA5 anomalies across the study period (α = 0.05).
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
e. 500-hPa geopotential height
Figure 7 shows the mean bias in the 500-hPa geopotential height field at 0 h (analysis, first column) and 48 h (second column), and the correlation between these biases (third column). All models exhibit an initial bias in 500-hPa geopotential height over portions of East Antarctica (Fig. 7). These biases persist to varying degree in all models except GDPS, in which positive oceanic biases present in all models at 48 h prevail. While these biases exist, it is important to note that they are seldom significant with respect to ERA5 anomalies over the period (see Fig. 7q for the sole exception), and only weakly correlated over land in some models.
500-hPa geopotential height mean bias at (left) 0 h (analysis) and (center) 48 h, and (right) the correlation coefficient between 0- and 48-h biases. Stippling (if illustrated) indicates significance from ERA5 anomalies across the study period (α = 0.05).
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
500-hPa geopotential height mean bias at (left) 0 h (analysis) and (center) 48 h, and (right) the correlation coefficient between 0- and 48-h biases. Stippling (if illustrated) indicates significance from ERA5 anomalies across the study period (α = 0.05).
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
500-hPa geopotential height mean bias at (left) 0 h (analysis) and (center) 48 h, and (right) the correlation coefficient between 0- and 48-h biases. Stippling (if illustrated) indicates significance from ERA5 anomalies across the study period (α = 0.05).
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
f. Parameter correlations
While we have examined a range of meteorological variables, there is also a strong interest to understand how forecast values and forecast biases correlate in space and time between pairs of study parameters and over different gridcell classifications (surface types). Figure 8 shows the Pearson’s R correlation coefficient between the 48-h forecast values (y axis) and the forecast bias values (x axis) between each pair of parameters discretized by cell classification. The purpose of this panel is to observe relationships between study parameters with the addition of surface fluxes in order to understand how they vary together under different cell classifications and whether forecast values and biases exhibit different correlative characteristics. Correlations are discussed using descriptors from Table 2, in terms of forecast correlation (where the values of each forecast parameter at a given coordinate correlate in time and space) and forecast bias correlation {where the difference between the modeled parameter and the reference analysis correlate in time and space [Eq. (7)]}.
Correlations between the 48-h forecast values (y axis) and the forecast bias values (x axis) between each pair of variables discretized by cell classification. Dotted and dashed lines indicate r = 0.5 and 0.7, respectively. Note, some models do not appear on all panels as one or both of the variables were not available in the model output to compute all correlations.
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
Correlations between the 48-h forecast values (y axis) and the forecast bias values (x axis) between each pair of variables discretized by cell classification. Dotted and dashed lines indicate r = 0.5 and 0.7, respectively. Note, some models do not appear on all panels as one or both of the variables were not available in the model output to compute all correlations.
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
Correlations between the 48-h forecast values (y axis) and the forecast bias values (x axis) between each pair of variables discretized by cell classification. Dotted and dashed lines indicate r = 0.5 and 0.7, respectively. Note, some models do not appear on all panels as one or both of the variables were not available in the model output to compute all correlations.
Citation: Weather and Forecasting 37, 6; 10.1175/WAF-D-21-0088.1
Correlation nomenclature used in this analysis.
The 2-m temperature and surface pressure forecasts are strongly correlated over land, with weak anticorrelation in biases in all models except ARPEGE-SH, which is considered moderate (Fig. 8a); this correlation in biases is reduced in ARPEGE-SH-GELATO. There is weak-to-moderate correlation between the parameters over ocean cells (and coastal cells in GDPS) and weak correlation elsewhere. The 2-m temperature and wind speed are weakly correlated in both forecast and forecast bias; however, ARPEGE-SH and ARPEGE-SH-GELATO show slightly higher correlation (albeit still weak) than other model in forecast error between the parameters, with the coupled variant showing lower correlation than the uncoupled (Fig. 8b). Forecast of wind speed over low concentration sea ice while weak is similarly correlated with forecast biases over the same cell classification. Forecasts of 2-m temperature and 500-hPa geopotential height (Fig. 8d) are strongly moderate-to-strongly correlated over the open ocean in all models, with ACCESS-G and GDPS showing similar correlation at the coast. Surface temperatures are moderate-to-strongly correlated with 500-hPa geopotential height over low-concentration sea ice in ACCESS-G, PolarWRF, ARPEGE-SH-GELATO, and ECMWF. Forecasts of 2-m temperature and sensible heat flux (Fig. 8g) are moderate-to-strongly correlated over low concentration sea ice, with weak correlation between the two parameters over the ocean in multiple models; in this example low concentration sea ice classifications yield stronger correlations than that of high concentration. Surface temperatures and latent heat flux (Fig. 8k) are weak-to-moderately correlated over low concentration sea ice and moderately anticorrelated over land. Again, low concentration sea is yields stronger correlation to that of high concentration sea ice. The difference between ARPEGE-SH and ARPEGE-SH-GELATO is once again highlighted forecast bias correlation with the coupled model yielding lower bias correlation than the uncoupled.
Surface pressure and wind speed forecasts (Fig. 8c) are weakly anticorrelated over both classifications of sea ice in all models; additionally, there is negligible correlation between forecasts and forecast biases in all models and cell classifications. There is strong correlation between surface pressure and 500-hPa geopotential height forecasts over open ocean, coastal, and sea ice cells in all models (Fig. 8e); however, correlations in their oceanic forecast biases are negligible unless the spatial scale is sufficiently coarsened (not shown). Surface pressure and latent heat flux are moderately anticorrelated in all models over land (Fig. 8l).
Wind speed and sensible heat flux forecasts are weak-to-moderately correlated over land in all models, with ARPEGE-SH forecast biases also weakly correlated between the parameters, a characteristic not shared by its coupled counterpart (Fig. 8i). All models and cell classifications show a weak anticorrelation between wind speed and latent heat flux (Fig. 8m), with a tendency for high concentration sea ice cells to yield stronger (albeit still weak) bias correlation than low concentration sea ice cells. The 500-hPa geopotential height and sensible heat flux forecasts (Fig. 8j) are weakly correlated in all models and classifications; however, low concentration and oceanic cells yield slightly stronger correlation between the parameters. This behavior is similar with regards to latent heat flux (Fig. 8n), with low concentration sea ice again showing slightly higher correlation between parameters, though still very weak in nature.
4. Discussion
Numerical weather prediction in general is considered to be an initial value problem, whereby small perturbations in a model’s initial conditions can rapidly spread through nonlinear processes to the entire model domain (Ancell et al. 2018); however, the problem can also be regarded as one of model dynamics and physics which is significantly more complicated than simply starting a forecast from more realistic initial conditions. Hence, Antarctic NWP is particularly challenging given the uncertainties surrounding sparse observations available for assimilation to form a set of initial conditions, as well as an incomplete understanding of the governing physical processes of the polar regions. While the YOPP austral summer SOP has provided a wealth of additional observations available for data assimilation and model verification, large regions of the continental interior, coastline, sea ice region and surrounding ocean remain without any observational record whatsoever (Schroeter et al. 2019). This places a stark contrast between Antarctica and more densely observed regions at lower latitudes and is an area subject to current discussion for the future needs of high southern latitude observations (see Pope et al. 2017). It suffices that model forecasts would typically diverge from the reference analysis at progressively longer lead times, particularly for those models that operate in an uncoupled, atmosphere-only capacity and with limited polar configuration in this model cohort. It is important to note that some forecast biases actually decreased over time depending on the model and parameter; however, this study is primarily focused on systemic bias biases and growing bias modes to identify opportunities for future model improvement.
a. Surface temperatures
The majority of the models evaluated in this study exhibit a warm analysis bias over land which can grow to as much as 3.5 K by 48 h over the continent when considering the range (SDE) of biases values across models (Figs. 2a, 3a, and 4). These biases are taken from the 0000 UTC forecast and as such do not capture the daily average bias or full diurnal temperature variation, which may be up to ±10°C in the continental interior. The Antarctic continent is known to exhibit a strong temperature inversion at the surface, whereby air over the frozen landmass may be substantially cooler than the air aloft (Hines and Bromwich 2008). The surface air therefore experiences negative buoyancy which contributes to the generation of katabatic wind (Connolley 1996). The implications of a positive (warm) surface temperature bias from excess entrainment of the warm air from above, should in theory result in a comparable negative (slow) bias in wind speed and vice versa, which is shown in Figs. 4 and 6 to varying degree between models. Curiously, despite a warm biased land surface temperature analysis and forecast in PolarWRF (Figs. 4d,e), wind speeds are only slower at analysis, with a positive (fast) wind speed bias developing at the coast at 48 h and with bias growth exceeding 2.5 m s−1 (Figs. 6d–f). Nigro et al. (2017) has suggested that the cold temperature bias in PolarWRF could be the result of a radiative imbalance in the model. While the Nigro et al. (2017) study applies to the flat Ross Ice Shelf, if this imbalance holds elsewhere excessively cold temperature biases on sloping coastal terrain may be responsible, at least in part, for strengthening gravity-driven katabatic winds around the continental edge (Figs. 6a–f). The (weak) moderate positive correlations between surface temperatures and sensible (latent) heat fluxes over low concentration sea ice exemplifies the need for quality sea ice consideration (be it coupling or otherwise) in a model and the biases that may be expected along the sea ice edge. Evidence of this can be seen in Fig. 4, whereby the only model of this cohort with a prescribed sea ice field, ACCESS-G, has widespread correlation between analysis biases and those at 48 h. In addition, the comparison between ARPEGE-SH-GELATO and ARPEGE-SH (two models differing only in coupling) shows that not only can sea ice coupling reduce forecast bias going forward (Fig. 2a), but the correlation between analysis and forecast error is also reduced (Figs. 4i,l). As 2-m temperature forecasts are generally well represented over open ocean in the majority of models examined (Fig. 2a), these results could imply that temperature inversions over the frozen landmass may not be adequately captured in several models, irrespective of forecast lead, and should be investigated using site observations and models with sufficient vertical resolution at the surface, which is beyond the scope of this study. The 2-m temperature biases are largely negative over sea ice (Fig. 2a) for multiple models, which could be attributed to the exposure of a warmer ocean due to the retreat of a comparatively cooler overlying sea ice layer, emphasizing the importance of atmosphere/sea ice coupling to avoid sizable biases in regions of sea ice advance and retreat (see Jung et al. 2016).
Surface temperature forecasts are at least moderately correlated with surface pressure over land, at the coast and over the open ocean in multiple models (Fig. 8a); however, forecast biases for these parameters are not convincingly correlated to suggest that an improvement in the predictability of one parameter would necessarily improve the other (Fig. 8a). Rather, the physical relationship between the parameters is made apparent. Further examination is required in order to fully understand the full nature of these correlations.
Surface temperatures off-continent, while negatively biased in several models out to 48hrs, are constrained most effectively through atmosphere–ocean–sea ice coupling. This can be observed in Fig. 4, whereby, coupled models (i.e., ARPEGE-SH-GELATO, ECMWF-IFS, and GDPS) exhibit smaller bias growth over ocean and sea ice than the remaining uncoupled models (noting the strong seasonality in ACCESS-G which drives mean error toward 0 across the study period). The sea ice edge is responsible for large contrasts in surface turbulent heat fluxes between sea ice and open ocean; however, high spatiotemporal resolution is required to adequately represent what is a rapidly changing surface environment on subdaily time scales (Jung et al. 2016). There are interesting correlations between forecast surface temperatures, surface winds and surface turbulent heat fluxes examined in this study, with weaker correlation in forecast bias for coupled models in these same variables over sea ice (Figs. 8g,i,k,m). Together, these results suggest that, by coupling atmosphere, ocean and sea ice models, surface turbulent heat flux biases at the sea ice edge are reduced, and that biases in surface forecasts of temperatures and winds can be de-correlated with biases in surface turbulent heat fluxes, potentially improving predictive skill in these regions. This is clear evidence that the addition of sea ice as a coupled model indeed improves predictive skill; however, differences in the correlative behaviors of surface temperatures and fluxes are clearly model specific.
Sea ice may experience considerable change during the course of a forecast due to physical processes occurring on subdaily time scales such as growth, melt and divergence at the sea ice edge (Handcock and Raphael 2020), which would not only alter the partitioning of sensible-to-latent flux via the Bowen ratio (thus rendering one term dominant over the other), but also changes in surface albedo and accumulated precipitation, the latter of which substantially alters the thermal properties of the underlying sea ice (Maykut and Untersteiner 1971). Figure 2a shows a common negative bias over sea ice among the models examined, with models forecasting surface temperatures colder than the reference analysis over both classifications of sea ice. An implication of these biases is an amplifying feedback mechanism, where cooler surface conditions delay the onset of summer sea ice retreat (Stammerjohn et al. 2012, 2008), which inhibits solar heating of the ocean and thus reduces the ice-free season and ocean–atmosphere heat and momentum exchange (Maksym et al. 2012). A runaway scenario is constrained through more frequent data assimilation (to realign the model with observations) or by coupling atmosphere and sea ice together. Figure 2a shows the positive impact of coupling; ARPEGE-SH-GELATO, ECMWF-IFS, and GDPS are three coupled systems, all of which have low surface temperature mean bias over low-concentration sea ice.
Coastal surface temperatures are cold biased in all models except ACCESS-G, whereby the positive surface temperature bias identified by Schroeter et al. (2019) is persistent throughout the forecast for the model and is associated with slower modeled winds at the coastal margins (Figs. 6a–c). Of the models exhibiting cold surface temperature biases, the resultant positive (fast) coastal wind bias is only notable in PolarWRF (Figs. 6d–f), with only localized regions of fast wind biases in other models.
The 2-m temperature forecasts over land are warm-biased at analysis and grow for the majority of the models examined with a warming (upward) trend as forecast length increases (Fig. 2a). When these biases are viewed in concert with SDE (Fig. 3a), they reach an order of up to 5 K at 48 h, depending on the model.
There is a moderate negative correlation between 2-m temperature and latent heat flux over land cells in all models, with weak bias correlation present in all models except ECMWF-IFS, suggesting that as latent heat exchange reduces, temperature increases due to the reduced availability of moist maritime air, and indicating other processes may have greater association with modeled surface temperatures in these models.
b. Surface pressure
Surface pressure is generally well represented in all models over ocean and both concentration classifications of sea ice (Fig. 2b); however, the range (SDE) of forecast biases is consistently higher at the coastal margins. This could be attributed to issues of topographic resolution between each model and the reference analysis (with lower resolution, though regridded in this study), whereby sharp changes in elevation from sea level may be substantial over the course of a model grid cell and thus may be subject to simple atomic bias resulting from gridcell discretization (i.e., values may be accurate but are off by a single grid cell); or through the passage of frequent transient cyclonic activity which contributes to wide variation in atmospheric pressure along the coast (Parish and Cassano 2003). The inability of a model to simulate katabatic winds at the coastal margins should also have an effect on surface pressures due to reduced drainage from the continental interior; however, there is little correlation between the parameters (both forecast and forecast bias) observed in this study, suggesting that these complex associations at resolutions higher than was assessed in this study. It is also important to note that most Antarctic observation sites are physically located at the coast near to sea level for reasons of accessibility; therefore, this large spread in the SDE may only be overcome through increased density of observations.
Surface pressure is strongly correlated with 500 hPa in all models over ocean and sea ice, which is consistent with Schroeter et al. (2019), who observed that surface pressure biases in ACCESS-G translate to a comparable displacement of the 500-hPa isobaric surface which was most prominent off-continent. However, the correlation between biases is only apparent at larger spatial scales (i.e., lower resolution; see supplementary Fig. 1e in the online supplemental material). The correlation in forecast biases between these two parameters over sea ice and ocean illustrates this relationship and warrants further investigation into the sensitivity of the 500-hPa isobaric surface to surface pressure biases (and vice versa). The implications of this relationship are of concern to operational Antarctic weather forecasting, as biases in the surface pressure and geopotential height fields can alter both the position and wind speeds of predicted transient cyclones, and particularly mesoscale cyclones, in the region due to the positioning of cyclone-generating polar lows brought about by katabatic wind activity from cold air advection (Carrasco and Bromwich 1993) and the potential vorticity (and thus rotational speed) of the atmospheric column.
Sea ice cover influences heat fluxes at the air–sea interface, as well as the absorption of shortwave radiation at the sea surface, processes which directly impact the variability of surface pressure and temperatures over the ice (Yuan and Li 2008). Though there is low correlation between the forecast values of surface pressure to sensible heat flux over sea ice, biases in surface pressure and sensible and latent heat flux are more strongly correlated over sea ice than other cell classifications for all models (supplementary Figs. 1h,l), with the surface pressure/latent heat flux bias correlations being the larger of the two flux components. The strength and sign of these bias correlations are clearly model dependent; however, the diversity of correlations to surface turbulent heat flux biases over the marginal ice zone (MIZ) where atmosphere, ocean, and sea processes are closely interrelated (Vichi et al. 2019) provides additional evidence of the sensitivity of surface parameters to biases in surface turbulent heat fluxes. One exemplar to the interconnected nature of the Antarctic system in this area is mesoscale cyclones, whereby boundary layer processes provide energy through turbulent surface heat fluxes and baroclinic instability to assist in and sustain cyclogenesis (Carrasco and Bromwich 1993; Uotila et al. 2011). Further to this, cyclones have been shown to both modify and be modified by surface heat fluxes (Bracegirdle and Kolstad 2010; Papritz et al. 2015; Uotila et al. 2011; Vichi et al. 2019).
Land biases in surface pressure exhibit a curious behavior in several models; in that the ME and SDE follow upward trends with regards to forecast length (Fig. 2b), but vacillate substantially between each of the forecast steps examined. One explanation of this is surface winds biases over the Antarctic interior, for which the majority of models produce slower modeled winds and, in many cases, become slower over time (Fig. 6). Antarctic interior winds are thought to be more closely associated with pressure gradient force due to radiational cooling and the comparatively more gentle slope of the terrain (Turner et al. 2009; Van den Broeke and Van Lipzig 2003). Given the low bias correlation between surface pressure and winds over land (Fig. 8c), this association is proven to be more sophisticated than can be quantified in Fig. 8 and further examination is required to form a better understanding of the processes involved.
c. 10-m wind speed
Surface wind speeds are well constrained to within a mean bias of ±0.5 m s−1 in all models over open ocean (Fig. 2c); however, an SDE analysis of these biases (Fig. 3c) indicates that these biases may indeed be fourfold out to 48 h. Figure 6 provides insight into the nature of surface winds in the marginal ice zone (MIZ) at larger spatial scales and the impact of a prescribed sea ice component on model performance. ACCESS-G (the only model in this cohort with a binary slab sea ice component), features not only slow modeled winds over sea ice, but also a strong correlation between biases at analysis and 48 h at the sea ice edge. This suggests that coupling can assist in de-correlating biases in wind speed over low concentration sea ice; however, it is important to note that this alone is insufficient to improve predictability in these regions and that further optimization through increased resolution, other polar enhancements, or the inclusion of an active ocean could improve upon low wind speed performance over sea ice.
Three of the remaining models of this cohort (ARPEGE-SH, ARPEGE-SH-GELATO, and GDPS) all feature slow modeled winds over the Antarctic continent (Fig. 6), which is likely associated with positive surface temperature biases that inhibit katabatic outflow. Conversely, PolarWRF develops a fast wind bias only at the coast, which becomes more positively biased as the forecast progresses (Figs. 6d–f). Given the downslope nature of these biases, they are likely dominated by the meridional component of wind, associated with cold surface temperature biases at the very edge of the continent (Figs. 2c, 3c, and 4d,e), the radiative imbalance suggested by Nigro et al. (2017), or some combination thereof.
d. 500-hPa geopotential height
The 500-hPa geopotential height fields are typically forecast too high over ocean the ocean (Figs. 2d and 3d) and over large regions of the continent at 48 h (Fig. 7); however, these biases of around 20 m are insignificant with regards to the reference analysis. There is a spatial discontinuity associated with the polar vortex surrounding Antarctica, highlighting the insular nature of the Antarctic atmosphere. The high correlations in all models between 500-hPa geopotential height and surface temperature forecasts over the ocean are an example of the meridional profile approaching the pole, where the 500-hPa surface is higher over the comparatively warmer ocean than it is over the frozen Antarctic landmass, rather than a strong association between the parameters (Fig. 8d). The 500-hPa geopotential height surface is guaranteed to be above the entire Antarctic landmass, where boundary layer effects are considered to be less influential (Bracegirdle and Marshall 2012). This holds for all but one of the surface parameters examined in this study, surface pressure. Surface pressure and 500-hPa geopotential height forecasts are shown to be strongly correlated in all models over ocean, coastal, and sea ice cell classifications (Fig. 8e); however, the correlation between their biases is unobservable at the high, native resolution of each of the models examined. A previous iteration of this study at 1° resolution found that biases between these two parameters were also highly correlated off-continent (supplementary Fig. 1e). This suggests that the vertical communication of surface pressure biases to the midtroposphere over the ocean found in ACCESS-G by Schroeter et al. (2019) may in fact be a systemic bias across multiple models, but that it occurs at meteorological scales larger than can be captured at higher model resolutions. In short, an elevated or depressed 500-hPa geopotential height surface over ocean and sea ice in the Antarctic can be associated with surface pressure biases below and vice versa due to the expansion or contraction of the atmospheric column, but that sufficient spatial scale is required to observe this association.
5. Conclusions
This study has evaluated six NWP models (coupled and uncoupled) in operational use over the Antarctic during the Year of Polar Prediction special observing period between 16 November 2018 and 15 February 2019. Most models are warm biased at the surface over land at the beginning of a forecast (analysis), with subtle negative biases around the coast and over the sea ice also evident. These land biases at analysis grow to around 3.5 K by 48 h in several of the models evaluated and the positive effect of coupling has been observed through the high correlation in surface temperature biases at analysis and 48 h over sea ice in uncoupled models. This bias growth is further evidence of the sensitivity of Antarctic NWP to initial conditions and highlighting the importance of model coupling and frequent, well-distributed and accurate observations; such is the goal of the special observing periods of the Year of Polar Prediction.
Surface pressure forecast biases were most variable at the coast and over ice shelves, while the forecasts themselves were found to be positively correlated with surface temperature outside of sea ice grid cells. This correlation is likely a reflection of the column dynamics or the associated surface pressure changes expected from a change in surface temperatures, varying regionally. There is a strong correlation between surface pressure and 500-hPa geopotential height, which suggests that pressure biases at the surface may be communicated vertically through to at least the midtroposphere in multiple models through an expansion or contraction of the atmospheric column above, resulting in an elevation or depression of the geopotential height surface; however, this association requires spatial scale larger than a model grid cell to observe.
Wind speed forecasts are well constrained over the ocean in all models; however, limited-area and higher-resolution models typically outperform global models over sea ice, where the latter typically forecast slower winds right up to the coast (ACCESS-G). However, differences between ERA5 and modeled wind speeds grow more rapidly at higher resolutions, which is consistent with studies in the Arctic (Køltzow et al. 2019).
This study has highlighted the association between multiple modeled forecast parameters and turbulent surface heat fluxes, whereby forecast and forecast bias correlations have been shown to yield the greatest diversity and complexity among the study parameters. These associations vary between model, parameter and cell classification; however, the quantification of these associations is beyond the scope of this study. Future studies should seek to characterize these associations so as to quantify the sensitivity of the Antarctic system to perturbations in surface turbulent heat fluxes and their influence elsewhere in the Antarctic system. While the number of models examined here is limited, there is distinct evidence that even at these relatively short periods, the inclusion of fully coupled sea ice and ocean components in these models compared with the uncoupled systems leads to lower forecast bias and lower correlations between forecast bias across variables.
Acknowledgments.
The authors wish to thank Sandrine Edouard from ECCC and Eric Bazile from Météo-France (CNRM-UMR3589) for their assistance in acquiring model output and documentation for GDPS and ARPEGE-SH, as well as Kirstin Werner, David Bromwich, and the generous members of the YOPP community for making their data available for analysis. The authors would also like to thank the three anonymous reviewers for their thorough critique of this study. This research was undertaken with the assistance of resources and services from the National Computational Infrastructure (NCI), which is supported by the Australian Government, and by the NeCTAR Research Cloud; a collaborative Australian research platform supported by the National Collaborative Research Infrastructure Strategy. This is a contribution to the Year of Polar Prediction (YOPP), a flagship activity of the Polar Prediction Project (PPP), initiated by the World Weather Research Programme (WWRP) of the World Meteorological Organisation (WMO). We acknowledge the WMO WWRP for its role in coordinating this international research activity. This project received funding from the Australian Government as part of the Antarctic Science Collaboration Initiative (ASCI000002).
Data availability statement.
Model output used in this study was acquired through each modeling center or from federated data portals: ACCESS-G APS2 via the NCI GeoNetwork Catalogue (https://geonetwork.nci.org.au), PolarWRF (AMPS) via the NCAR Climate Data Gateway (https://www.earthsystemgrid.org), and ECMWF-IFS via the ECMWF Public Datasets Collection (https://apps.ecmwf.int/datasets). ARPEGE-SH, ARPEGE-SH-GELATO (Météo-France), and GDPS (ECCC) data can be acquired by contacting corresponding authors cited in the literature referenced in this study. ERA5 data are available from the Copernicus Climate Data Store (https://cds.climate.copernicus.eu).
REFERENCES
Albergel, C., E. Dutra, S. Munier, J.-C. Calvet, J. Munoz-Sabater, P. de Rosnay, and G. Balsamo, 2018: ERA-5 and ERA-Interim driven ISBA land surface model simulations: Which one performs better? Hydrol. Earth Syst. Sci., 22, 3515–3532, https://doi.org/10.5194/hess-22-3515-2018.
Ancell, B. C., A. Bogusz, M. J. Lauridsen, and C. J. Nauert, 2018: Seeding chaos: The dire consequences of numerical noise in NWP perturbation experiments. Bull. Amer. Meteor. Soc., 99, 615–628, https://doi.org/10.1175/BAMS-D-17-0129.1.
Australian Bureau of Meteorology, 2016: APS2 upgrade to the ACCESS-G Numerical Weather Prediction System. BNOC Operations Bulletin 105, Australian Bureau of Meteorology, 32 pp., http://www.bom.gov.au/australia/charts/bulletins/APOB105.pdf.
Bauer, P., I. Sandu, L. Magnusson, R. Mladek, and M. Fuentes, 2020: ECMWF global coupled atmosphere, ocean and sea-ice dataset for the Year of Polar Prediction 2017–2020. Sci. Data, 7, 427, https://doi.org/10.1038/s41597-020-00765-y.
Bazile, E., N. Azouz, A. Napoly, and C. Loo, 2020: Impact of the 1D sea-ice model GELATO in the global model ARPEGE. WCRP Rep. 12/2020 (Research Activities in Earth System Modelling). WMO, 2 pp., http://bluebook.meteoinfo.ru/uploads/2020/docs/06_Bazile_Eric_sea_ice_model_in_ARPEGE.pdf.
Belmonte Rivas, M., and A. Stoffelen, 2019: Characterizing ERA-Interim and ERA5 surface wind biases using ASCAT. Ocean Sci., 15, 831–852, https://doi.org/10.5194/os-15-831-2019.
Bracegirdle, T. J., and E. W. Kolstad, 2010: Climatology and variability of Southern Hemisphere marine cold-air outbreaks. Tellus, 62, 202–208, https://doi.org/10.1111/j.1600-0870.2009.00431.x.
Bracegirdle, T. J., and G. J. Marshall, 2012: The reliability of Antarctic tropospheric pressure and temperature in the latest global reanalyses. J. Climate, 25, 7138–7146, https://doi.org/10.1175/JCLI-D-11-00685.1.
Bromwich, D. H., A. J. Monaghan, K. W. Manning, and J. G. Powers, 2005: Real-time forecasting for the Antarctic: An evaluation of the Antarctic Mesoscale Prediction System (AMPS). Mon. Wea. Rev., 133, 579–603, https://doi.org/10.1175/MWR-2881.1.
Bromwich, D. H., and Coauthors, 2020: The year of polar prediction in the Southern Hemisphere (YOPP-SH). Bull. Amer. Meteor. Soc., 101, E1653–E1676, https://doi.org/10.1175/BAMS-D-19-0255.1.
Carrasco, J. F., and D. H. Bromwich, 1993: Mesoscale cyclogenesis dynamics over the southwestern Ross Sea, Antarctica. J. Geophys. Res., 98, 12 973–12 995, https://doi.org/10.1029/92JD02821.
Connolley, W. M., 1996: The Antarctic temperature inversion. Int. J. Climatol., 16, 1333–1342, https://doi.org/10.1002/(SICI)1097-0088(199612)16:12<1333::AID-JOC96>3.0.CO;2-6.
Dong, X., Y. Wang, S. Hou, M. Ding, B. Yin, and Y. Zhang, 2020: Robustness of the recent global atmospheric reanalyses for Antarctic near-surface wind speed climatology. J. Climate, 33, 4027–4043, https://doi.org/10.1175/JCLI-D-19-0648.1.
Environment and Climate Change Canada, 2018: The Global Deterministic Prediction System (GDPS) version 6.1.0 of the Meteorological Service (MSC) of Canada. Environment and Climate Change Canada, 9 pp., https://collaboration.cmc.ec.gc.ca/cmc/cmoi/product_guide/docs/tech_specifications/tech_specifications_GDPS_6.1.0_e.pdf.
Grenfell, T. C., S. G. Warren, and P. C. Mullen, 1994: Reflection of solar radiation by the Antarctic snow surface at ultraviolet, visible, and near‐infrared wavelengths. J. Geophys. Res., 99, 18 669–18 684, https://doi.org/10.1029/94JD01484.
Hall, A., 2004: The role of surface albedo feedback in climate. J. Climate, 17, 1550–1568, https://doi.org/10.1175/1520-0442(2004)017<1550:TROSAF>2.0.CO;2.
Handcock, M. S., and M. N. Raphael, 2020: Modeling the annual cycle of daily Antarctic sea ice extent. Cryosphere, 14, 2159–2172, https://doi.org/10.5194/tc-14-2159-2020.
Hersbach, H., and Coauthors, 2020: The ERA5 global reanalysis. Quart. J. Roy. Meteor. Soc., 146, 1999–2049, https://doi.org/10.1002/qj.3803.
Hillebrand, F. L., U. F. Bremer, J. Arigony-Neto, C. N. da Rosa, C. W. Mendes, J. Costi, M. W. D. de Freitas, and F. Schardong, 2021: Comparison between atmospheric reanalysis models ERA5 and ERA-Interim at the North Antarctic Peninsula Region. Ann. Assoc. Amer. Geogr., 111, 1147–1159, https://doi.org/10.1080/24694452.2020.1807308.
Hines, K. M., and D. H. Bromwich, 2008: Development and testing of polar Weather Research and Forecasting (WRF) Model. Part I: Greenland ice sheet meteorology. Mon. Wea. Rev., 136, 1971–1989, https://doi.org/10.1175/2007MWR2112.1.
Hudson, S. R., and R. E. Brandt, 2005: A look at the surface-based temperature inversion on the Antarctic Plateau. J. Climate, 18, 1673–1696, https://doi.org/10.1175/JCLI3360.1.
Jung, T., and Coauthors, 2016: Advancing polar prediction capabilities on daily to seasonal time scales. Bull. Amer. Meteor. Soc., 97, 1631–1647, https://doi.org/10.1175/BAMS-D-14-00246.1.
Køltzow, M., B. Casati, E. Bazile, T. Haiden, and T. Valkonen, 2019: An NWP model intercomparison of surface weather parameters in the European Arctic during the year of polar prediction special observing period Northern Hemisphere 1. Wea. Forecasting, 34, 959–983, https://doi.org/10.1175/WAF-D-19-0003.1.
Kuipers Munneke, P., M. R. van den Broeke, J. T. M. Lenaerts, M. G. Flanner, A. S. Gardner, and W. J. van de Berg, 2011: A new albedo parameterization for use in climate models over the Antarctic ice sheet. J. Geophys. Res., 116, D05114, https://doi.org/10.1029/2010JD015113.
Maksym, T., S. E. Stammerjohn, S. Ackley, and R. Massom, 2012: Antarctic sea ice—A polar opposite? Oceanography, 25, 140–151, https://doi.org/10.5670/oceanog.2012.88.
Martens, B., D. L. Schumacher, H. Wouters, J. Muñoz-Sabater, N. E. C. Verhoest, and D. G. Miralles, 2020: Evaluating the land-surface energy partitioning in ERA5. Geosci. Model Dev., 13, 4159–4181, https://doi.org/10.5194/gmd-13-4159-2020.
Martin, P. J., and D. A. Peel, 1978: The spatial distribution of 10 m temperatures in the Antarctic Peninsula. J. Glaciol., 20, 311–317, https://doi.org/10.1017/S0022143000013861.
Maykut, G. A., 1978: Energy exchange over young sea ice in the central Arctic. J. Geophys. Res., 83, 3646–3658, https://doi.org/10.1029/JC083iC07p03646.
Maykut, G. A., and N. Untersteiner, 1971: Some results from a time‐dependent thermodynamic model of sea ice. J. Geophys. Res., 76, 1550–1575, https://doi.org/10.1029/JC076i006p01550.
Nigro, M. A., J. J. Cassano, J. Wille, D. H. Bromwich, and M. A. Lazzara, 2017: A self-organizing-map-based evaluation of the Antarctic Mesoscale Prediction System using observations from a 30-m instrumented tower on the Ross Ice Shelf, Antarctica. Wea. Forecasting, 32, 223–242, https://doi.org/10.1175/WAF-D-16-0084.1.
Pailleux, J., and Coauthors, 2015: Les 25 ans du système de prévision numérique du temps IFS/Arpège. Meteorologie, 89, 18–27, https://doi.org/10.4267/2042/56594.
Papritz, L., S. Pfahl, H. Sodemann, and H. Wernli, 2015: A climatology of cold air outbreaks and their impact on air–sea heat fluxes in the high-latitude South Pacific. J. Climate, 28, 342–364, https://doi.org/10.1175/JCLI-D-14-00482.1.
Parish, T. R., and D. H. Bromwich, 1987: The surface windfield over the Antarctic ice sheets. Nature, 328, 51–54, https://doi.org/10.1038/328051a0.
Parish, T. R., and J. J. Cassano, 2003: The role of katabatic winds on the Antarctic surface wind regime. Mon. Wea. Rev., 131, 317–333, https://doi.org/10.1175/1520-0493(2003)131<0317:TROKWO>2.0.CO;2.
Pope, A., P. Wagner, R. Johnson, J. D. Shutler, J. Baeseman, and L. Newman, 2017: Community review of Southern Ocean satellite data needs. Antarct. Sci., 29, 97–138, https://doi.org/10.1017/S0954102016000390.
Powers, J. G., K. W. Manning, D. H. Bromwich, J. J. Cassano, and A. M. Cayette, 2012: A decade of Antarctic science support through AMPS. Bull. Amer. Meteor. Soc., 93, 1699–1712, https://doi.org/10.1175/BAMS-D-11-00186.1.
Puri, K., and Coauthors, 2013: Implementation of the initial ACCESS numerical weather prediction system. Aust. Meteor. Oceanogr. J., 63, 265–284, https://doi.org/10.22499/2.6302.001.
Schroeter, B. J. E., P. Reid, N. L. Bindoff, and K. Michael, 2019: Antarctic verification of the Australian numerical weather prediction model. Wea. Forecasting, 34, 1081–1096, https://doi.org/10.1175/WAF-D-18-0171.1.
Sheridan, P., S. Smith, A. Brown, and S. Vosper, 2010: A simple height-based correction for temperature downscaling in complex terrain. Meteor. Appl., 17, 329–339, https://doi.org/10.1002/met.177.
Soldatenko, S., C. Tingwell, P. Steinle, and B. A. Kelly-Gerreyn, 2018: Assessing the impact of surface and upper-air observations on the forecast skill of the ACCESS numerical weather prediction model over Australia. Atmosphere, 9, 23, https://doi.org/10.3390/atmos9010023.
Stammerjohn, S. E., D. Martinson, R. Smith, X. Yuan, and D. Rind, 2008: Trends in Antarctic annual sea ice retreat and advance and their relation to El Niño–Southern Oscillation and Southern Annular Mode variability. J. Geophys. Res., 113, C03S90, https://doi.org/10.1029/2007JC004269.
Stammerjohn, S. E., R. Massom, D. Rind, and D. Martinson, 2012: Regions of rapid sea ice change: An inter‐hemispheric seasonal comparison. Geophys. Res. Lett., 39, L06501, https://doi.org/10.1029/2012GL050874.
Tetzner, D., E. Thomas, and C. Allen, 2019a: A validation of ERA5 reanalysis data in the southern Antarctic Peninsula–Ellsworth Land region, and its implications for ice core studies. Geosciences, 9, 289, https://doi.org/10.3390/geosciences9070289.
Turner, J., S. N. Chenoli, A. abu Samah, G. Marshall, T. Phillips, and A. Orr, 2009: Strong wind events in the Antarctic. J. Geophys. Res., 114, D18103, https://doi.org/10.1029/2008JD011642.
Uotila, P., T. Vihma, A. B. Pezza, I. Simmonds, K. Keay, and A. H. Lynch, 2011: Relationships between Antarctic cyclones and surface conditions as derived from high-resolution numerical weather prediction data. J. Geophys. Res., 116, D07109, https://doi.org/10.1029/2010JD015358.
Van den Broeke, M. R., and N. Van Lipzig, 2003: Factors controlling the near-surface wind field in Antarctica. Mon. Wea. Rev., 131, 733–743, https://doi.org/10.1175/1520-0493(2003)131<0733:FCTNSW>2.0.CO;2.
Vichi, M., and Coauthors, 2019: Effects of an explosive polar cyclone crossing the Antarctic marginal ice zone. Geophys. Res. Lett., 46, 5948–5958, https://doi.org/10.1029/2019GL082457.
Wang, C., R. M. Graham, K. Wang, S. Gerland, and M. A. Granskog, 2019: Comparison of ERA5 and ERA-Interim near-surface air temperature, snowfall and precipitation over Arctic sea ice: Effects on sea ice thermodynamics and evolution. Cryosphere, 13, 1661–1679, https://doi.org/10.5194/tc-13-1661-2019.
Yamanouchi, T., and S. Kawaguchi, 1984: Longwave radiation balance under a strong surface inversion in the Katabatic Wind Zone, Antarctica. J. Geophys. Res., 89, 11 771–11 778, https://doi.org/10.1029/JD089iD07p11771.
Yuan, X., and C. Li, 2008: Climate modes in southern high latitudes and their impacts on Antarctic sea ice. J. Geophys. Res., 113, C06S91, https://doi.org/10.1029/2006JC004067.
Zhu, J., A. Xie, X. Qin, Y. Wang, B. Xu, and Y. Wang, 2021: An assessment of ERA5 reanalysis for Antarctic near-surface air temperature. Atmosphere, 12, 217, https://doi.org/10.3390/atmos12020217.
