1. Introduction
The Antarctic observational network began during the 1957–58 International Geophysical Year with year-round staffed stations, and grew rapidly in the 1970s and 1980s with the development of automated weather stations (AWSs) that could withstand the harsh weather conditions and transmit data to polar-orbiting satellites (Bromwich et al. 2013a; Lazzara et al. 2012). The University of Wisconsin AWS network (run by the Antarctic Meteorological Research Center) was an integral part of this growth and is currently responsible for over half of the AWSs in Antarctica. Despite the evolving network improvements, coverage remains sparse with interior observations particularly lacking (Chapman and Walsh 2007; Lazzara et al. 2012). Different monitoring goals, improving technology, and inconsistency in the availability of funding, among other practical issues, have all influenced station placement in the network. Logistical challenges for installing and maintaining an observational network that spans a continent more than half the size of North America are numerous, which motivates research into optimal configurations of the network to maximize performance for a given monitoring goal (e.g., Huntley and Hakim 2010; Mauger et al. 2013). Here we assess the spatial coverage of the current Antarctica network as a precursor to a complete network design study.
Surface weather measurements in Antarctica are critical for supporting base operations, especially at McMurdo station, the main U.S. base in Antarctica (Steinhoff et al. 2008). To further support the U.S. Antarctic Program (USAP) operations at McMurdo and other stations, the Antarctic Mesoscale Prediction System (AMPS), a collaboration between the National Center for Atmospheric Research (NCAR) and The Ohio State University (OSU), began in 2000. The primary goals of the AMPS program are to provide synoptic and mesoscale model output to forecasters and to improve model development for the Antarctic (Powers et al. 2012). Beginning in 2008, AMPS has used the Weather Research and Forecasting (WRF; Skamarock et al. 2008) Model, with modifications for polar environments, to provide forecasts over Antarctica. Before 2008, AMPS used a polar-optimized version of the fifth-generation Pennsylvania State University–NCAR Mesoscale Model (MM5; Bromwich et al. 2001). AMPS uses a three-dimensional variational algorithm (3DVAR) for assimilation of radiosonde, surface, and satellite data, using the 0000 UTC Global Forecast System (GFS) analysis as a background estimate.
Studies have shown good agreement between AMPS forecasts and observations (e.g., Nicolas and Bromwich 2011; Bromwich et al. 2005; Guo et al. 2003), and the AMPS data have been used to study specific weather phenomena (e.g., Nigro et al. 2012; Steinhoff et al. 2008) and climate (e.g., Seefeldt and Cassano 2012; Monaghan et al. 2005) over Antarctica. Though many of these studies were performed when AMPS employed Polar MM5, Bromwich et al. (2013b) found that the Polar WRF performed similarly or better than the Polar MM5. Here AMPS is utilized to evaluate the spatial coverage of the current observational network, focusing on temperature and pressure.
The remainder of this paper is organized as follows. AMPS data and surface observations are described in section 2, along with methods used for quality control and removal of the seasonal cycle. In section 3, we present results on the effectiveness of the current network at describing mean sea level pressure and 2-m temperature on different time scales using spatial correlation analysis and multilinear regression. Section 4 provides the conclusions and initial estimates of optimal station placement for different regions of the continent.
2. Data and methods
a. Antarctic Mesoscale Prediction System
This study uses the 15-km “continental grid” AMPS 0000 UTC analyses from 1 October 2008 to 31 October 2012 from NCAR, during which time the model grid is unchanged. Surface temperature exhibits a strong seasonal cycle that dominates estimates of spatial correlation. In contrast, a much less distinct, but documented seasonal cycle in surface pressure (Parish and Bromwich 1997) is found to have a negligible influence on the spatial correlations, and therefore is not removed. To remove the seasonal cycle in temperature, we first detrend the time series at each point, and then remove all Fourier harmonics with periods longer than 6 months. To address the dependence of results on season, due to changes in daily variability, two 6-month periods are defined as rough definitions of austral winter [April–September (AMJJAS)] and austral summer [October–March (ONDJFM)]. Shorter definitions of the seasons are also used (i.e., December–February, March–May, June–August, and September–November) in some cases, as indicated below. Unless otherwise noted, the terms winter and summer refer to the 6-month season definitions.
b. Observations
Raw, “real time” Antarctic surface weather observations, that include both U.S. and international AWSs and staffed stations, were obtained from NCAR for the period of study. These observations are used by the AMPS data assimilation system and are inputted to the WRF variational data assimilation (WRFDA) observation preprocessor. While more stations exist in Antarctica, this analysis is focused on the station data available to AMPS in real time.
1) Quality control
Basic quality control (QC) is performed on the daily observations as follows:
Remove obviously erroneous values that are not within the expected range of what is physically possible (e.g., pressure must be >0, pressure must be <1013 hPa if station height is above 350 m).
Remove consecutive daily observations of the same value that are indicative of measurement or transmitting error.
Remove outliers from the data.
Since the purpose of this work is to assess the utility of the current network for weather monitoring and forecasting, our analysis is focused on observational stations with a minimum of missing or erroneous data. Specifically, the quality-controlled observations were broken into two sets: those having valid observations for at least 90% of the record [i.e., the “best” coverage; hereafter complete data 90% (CD90)] and those having at least 75% of the record [i.e., “good” coverage; hereafter complete data 75% (CD75)]. These two sets are used in the analysis described below; Table 1 shows the number of temperature and pressure stations that were included in each set. Since the analysis focuses mostly on temperature, a map of stations included in the CD75 and CD90 set for temperature is shown in Fig. 1, as well as geographical labels for areas mentioned in this paper.
The number of stations for both temperature and pressure that were reporting for at least 75% of the 4-yr period (CD75) and for at least 90% of the period (CD90).
A map of Antarctica showing the temperature stations in the CD75 (all markers) and the CD90 (squares) subsets. The three sample stations used to represent different regions of the continent—Rothera Point, Theresa, and Vostok—are highlighted with a black square surrounding the marker. Geographical regions that are mentioned in the text are also denoted for reference.
Citation: Monthly Weather Review 142, 10; 10.1175/MWR-D-13-00401.1
2) Seasonal cycle
As with the AMPS data, we detrended each observation time series and removed the seasonal cycle for temperature by subtracting all harmonics with frequencies lower than semiannual; missing daily observations are replaced with zeroes. Although artificial zeroes are likely to only affect high frequencies, this introduces the possibility of a biased estimate of the power spectrum. To test this possibility, the harmonics were recomputed after replacing the missing values using linear interpolation. Differences between the two approaches are minor, and have little effect on the spatial correlations that are the focus of this study.
c. Correlation length scales
Spatial correlations were estimated by calculating correlations between a selected grid point and a subset of grid points within the continent. For ease of computation, the correlations were evaluated for every fifth grid point. Assuming the correlation decays exponentially with distance, the correlation length scale (CLS) is estimated using the e-folding distance of the correlation between grid points. Three steps are taken in order to obtain a robust estimate of the average correlation distance. 1) The distance is defined as twice the e-folding length of the squared correlation, which more closely approximates exponential decay. 2) The e-folding length is estimated using the average distance for all points for which the squared correlation is between e−0.9 and e−1.1. Tests indicate that the results are not sensitive to the choice of this interval (not shown). We chose to composite correlations near 1/e instead of using linear regression because the latter was influenced by departures from exponential decay at long distances. 3) The confidence limits are estimated using a bootstrap approach by iteratively resampling, with replacement, and recomputing the average correlation distance from each subsample.
d. Variance explained by the current network
To determine how well the current network of weather stations explains the temperature and pressure variance across the continent, a multilinear regression is performed. The AMPS grid points are used as proxies for the location of the station observations, and are regressed onto the AMPS time series at each grid point. We use the CD75 and CD90 stations as separate predictors, and perform separate calculations for daily, weekly, and monthly averaged time series. The variance explained is the squared correlation between the regression and the actual time series (e.g., Wilks 2011). This approach assumes that the AMPS data are a good proxy for actual observations. In the following section, we show that this is a good assumption.
3. Results
a. Spatial correlations for the current network
The goal in this section is to use the correlations between individual observing stations to 1) estimate the spatial correlation length scale for three regions of the continent directly from the observations and 2) compare the results with those for AMPS. Results show that the AMPS data agree well with the observations, which lends confidence to using AMPS for a more comprehensive analysis evaluating cross correlations over the entire continent. Since the observation sites have widely variable reporting frequencies, we distinguish and compare results for stations in the CD75 and CD90 subsets.
For the CD75 stations, correlations apply to each pair of the 45 temperature (Fig. 2) and 44 pressure stations (Fig. 3). Three reference stations were chosen as illustrative examples of the three regions of the continent: the Antarctic Peninsula (Rothera Point), West Antarctica (Theresa), and East Antarctica (Vostok). Correlations are plotted as a function of distance from each reference station.
Correlations in the daily temperature data ranked by distance between sites on (top) the Antarctic Peninsula (Rothera Point at 67.56°S, 68.13°W), (middle) West Antarctica (Theresa at 84.60°S, 115.82°W), and (bottom) East Antarctica (Vostok at 78.45°S, 106.87°W) and other station locations (in CD75) for the observations (red line) and the AMPS 0000 UTC analysis (black line). Correlations between each of the three sites are highlighted on the observation time series.
Citation: Monthly Weather Review 142, 10; 10.1175/MWR-D-13-00401.1
As in Fig. 2, but for surface pressure.
Citation: Monthly Weather Review 142, 10; 10.1175/MWR-D-13-00401.1
As expected, temperature (Fig. 2) and pressure (Fig. 3) correlations decrease with distance from the reference station. Surface pressure spatial correlations are higher than those for temperature for all three regions, indicating a greater regional coherence in daily pressure variations. For both variables the correlations decrease at a faster rate for the peninsula and a greater number of stations have essentially no correlation with the peninsula site, as was also found in King and Comiso (2003) using satellite measurements, suggesting that the peninsula is dynamically distinct from the rest of the continent. This is consistent with previous work indicating the enhanced influence of the southern annular mode (SAM) on the Antarctic Peninsula (Thompson et al. 2011), and the different seasonal timing of tropical influences in this region compared with West Antarctica (Ding and Steig 2013). Correlations between the West Antarctic (Theresa) and East Antarctic (Vostok) sites are the highest among the three representative sites for both temperature (0.20) and pressure (0.65).
Comparing results for correlations between the stations using the observations (Figs. 2 and 3; red line) and the corresponding grid points in AMPS (Figs. 2 and 3; black line) shows good agreement for both variables. In both cases, AMPS tends to slightly overestimate the correlations at distances within 1000 km of the sample sites. For the three reference sites, AMPS temperature correlations are biased slightly high for stations within 1000 km and biased slightly low for stations that are located farther away (Table 2). Overall, the very good agreement between the observations and AMPS correlations suggests that using AMPS as a proxy for station “locations” elsewhere in the continent is appropriate. Correlations between the observations and AMPS also compare closely for 12-, 24-, and 36-h AMPS forecast grids (not shown), suggesting that the close correspondence is not simply an artifact of assimilating observations. This point is reinforced by a separate calculation for observations that are not assimilated. We identify 12 stations that are assimilated into AMPS less than 1% of the 4-yr period and that are included in the CD75 set (Rothera Point is one such station). AMPS correlations for this subset are also in good agreement with observations (not shown).
The mean temperature correlation bias between AMPS and the observations for stations within 1000 km of the reference site and for stations 1000 km away from the reference site. An estimate of the uncertainty in the mean is shown in parentheses.
b. Correlation length scales using AMPS
Using a subsample of the AMPS data (sampling at one-fifth of the grid resolution), differences in temperature correlations between the three regions of Antarctica are examined further. For each region, the correlation between each grid point and every other grid point in the region is computed. This is repeated for all grid points in the region and binned by the distance between points, as displayed in Fig. 4. As previously described, the CLS is defined as the e-folding squared correlation decay distance for each region. Results show a longer average CLS in East Antarctica (1370 km; 95% confidence range: 1360–1380 km) when compared to West Antarctica (1110 km; 95% confidence range: 1090–1130 km) and the peninsula (840 km; 95% confidence range: 790–890 km). Note that the confidence limits represent the uncertainty in the average CLS estimated for each region, not the range of CLS within each region. The longer CLS in the high plateau of East Antarctica implies that fewer stations per unit area are needed in that area to capture daily variations in weather.
Box-and-whisker plots of the correlation ranked by distance for a subsample of AMPS grid points for (a) the Antarctic Peninsula, (b) West Antarctica, and (c) East Antarctica. The points are placed into bins that are 150 km in width. The box indicates the 25th, 50th, and 75th percentiles, and the whiskers enclose 99.3% of the distribution. Outliers are excluded. The open circles indicate the average correlation length scale for each region as defined as the e-folding length of decay in correlation with distance.
Citation: Monthly Weather Review 142, 10; 10.1175/MWR-D-13-00401.1
To determine the spatial and seasonal variability in the CLS, a similar calculation was performed for the whole continent using the AMPS data. An estimated CLS for each grid point is shown in Fig. 5 for austral summer and winter, separately. Again results show longer correlation length scales over most of East Antarctica, especially in summer where they range between about 1300 and 2100 km. Note that CLS is longer here than what is implied in Fig. 4 because the calculation includes points from the entire continent, not just those within each individual region. The results also show significant spatial variability, which implies that some observing locations provide more spatial information than others. For example, over West Antarctica, an area adjacent to the Ronne Ice Shelf has relatively long correlation length scales compared to the rest of the region. The shorter CLS in West Antarctica, and particularly the peninsula, suggest that more stations per unit area are needed in those regions to capture spatial variations in weather. Pine Island Bay, an area of increasing research interest because of the rapid climate-driven glaciological changes occurring there (e.g., Thoma et al. 2008; Jenkins et al. 2010; Steig et al. 2012), has a CLS that ranges between 1100 and 1300 km for the two seasons, which is short compared to those for most of East Antarctica, but relatively long compared to those in West Antarctica. A nearby station may be able to monitor weather for this location provided it was situated close enough.
Estimated fitted CLS (km) for (top) winter and (bottom) summer days using daily AMPS data.
Citation: Monthly Weather Review 142, 10; 10.1175/MWR-D-13-00401.1
In general, the CLS is longer in the summer, although this is not always the case. For the peninsula and the coastal area adjacent to the Weddell Sea, the CLSs are shorter during summer when compared to winter. This suggests that the temperatures across the peninsula do not vary uniformly during summer, which is consistent with an analysis by Ding and Steig (2013), which found that temperature variability in the western Antarctic Peninsula is independent from the eastern peninsula during summer (DJF), but not during the other three seasons. Figure 5 also reveals that the general patterns of the CLS are different for the two seasons, suggesting that the CLS are related to the seasonal weather patterns and are not entirely terrain driven.
It is important to note that even the short CLSs over Antarctica are relatively long; the scale in Fig. 5 begins at 440 km (or roughly about 4° of latitude). Considering the challenges that exist for installing and maintaining weather stations in the harsh Antarctic climate, relatively long CLSs have positive implications for enhancing the current network for weather and climate monitoring.
c. Variance explained by the current network
We transition now to assessing the pressure and temperature variance explained by the current observing network. The variance explained is calculated using multiple linear regression, where gridpoint AMPS data are regressed onto station locations for the 4-yr period. The calculation was performed using daily, weekly, and monthly time averages, as well as for austral winter and summer. Results for the CD75 and CD90 datasets are also compared to assess the influence of station availability.
The variance explained for temperature is shown in Fig. 6 for CD90 and in Fig. 7 for CD75. Comparing daily, weekly, and monthly time averages, it is evident that coverage improves substantially with increasing time average.1 Thus, fewer stations are needed to capture the temperature variability on monthly time scales. However, gaps in network coverage can be identified for daily and weekly time averages. It is interesting how little the station on the west side of the peninsula represents the east side of the peninsula in CD90 (Fig. 6); this result is consistent with previous work using spatial correlations from satellite data that illustrates the differing air mass influence divided by the steep terrain (King and Comiso 2003). Additionally, variability near the Ronne Ice Shelf in West Antarctica is not captured for daily or even weekly time averages for CD90. Even when moving to the subset that includes more stations, CD75, large areas of the continent remain undersampled on daily time scales, including Pine Island Bay and the coast extending to the Ross Ice Shelf, the Ronne Ice Shelf, and eastern Antarctica in the Indian Ocean sector. Conversely, variance in surface pressure is well explained (ranging from 0.8 to 1) throughout the continent on daily time scales even using the CD90 subset of stations (not shown).
Fraction of temperature variance explained for (top) daily, (middle) weekly, and (bottom) monthly time averages for (left) winter and (right) summer for the stations that reported 90% of the time (CD90) during the 4-yr period. Note that the color scales are different for each panel, with the colors representing a smaller range as the time average increases. The daily data range from 0 to 1, the weekly data averages from 0.3 to 1, and the monthly data averages from 0.7 to 1. The sample stations used in Figs. 2 and 3 are denoted by the square (Rothera Point) and the triangle (Vostok).
Citation: Monthly Weather Review 142, 10; 10.1175/MWR-D-13-00401.1
As in Fig. 6, but for only (top) daily and (bottom) weekly time averages and stations that are reported temperature 75% of the time (CD75). The daily data range from 0 to 1 and the weekly data averages from 0.7 to 1. The sample stations used in Figs. 2 and 3 are denoted by the square (Rothera Point), the larger circle (Theresa), and the triangle (Vostok). Note that the monthly averages are not included here because there were too few months to perform the multilinear regression using all CD75 stations (i.e., there were fewer months than stations, resulting in a rank deficient calculation).
Citation: Monthly Weather Review 142, 10; 10.1175/MWR-D-13-00401.1
Seasonal differences between the temperature variance explained in winter versus summer are also evident in Figs. 6 and 7. In general, the interior stations explain more variance over a larger area for temperature during summer, which can also be seen in the longer summer correlation length scales shown in Fig. 5. The South Pole station and Vostok (denoted by the triangle) are good examples of areas having larger spatial influence in summer. On the other hand, in general, the coastal stations explain more variance over a larger area for temperature during winter when compared to summer (Figs. 6 and 7). The variance explained along the coast is very locally constrained, likely due to winter storms that do not often penetrate inland beyond the steep coastal topography (Nicolas and Bromwich 2011). Three-month seasonal breakdowns (i.e., DJF, MAM, JJA, and SON) reveal similar seasonal differences between the interior and the coast; austral fall (MAM) and winter (JJA) look very much like AMJJAS while austral summer (DJF) looks very similar to ONDJFM (not shown). One exception concerns spring (SON), where the area of higher temperature variance explained is larger than in DJF and ONDJFM, even on the coast (not shown).
We hypothesize that the larger summer spatial influence of the interior stations is due to the lack of the persistent winter temperature inversion (e.g., Warren and Town 2011) during summer. A weaker or nonexistent inversion allows greater mixing and less influence of local surface-based features, yielding longer correlation length scales during summer. For coastal stations, we hypothesize that the larger spatial influence in winter is due to persistent katabatic winds (e.g., Parish and Bromwich 2007) that produce widespread downslope winds as air moves down complex terrain toward the coastline.
4. Discussion and conclusions
Because of spatial correlations in field variables, station measurements provide information over a region surrounding the location of the instrument. The spatial extent of that influence provides a measure of the effectiveness of the station network for monitoring weather and climate. This study evaluates the performance of the weather station network over Antarctica for monitoring surface temperature and pressure. It should be noted that more stations exist on the continent than are considered here; we considered only those stations in which data are available to AMPS in real time. For monitoring stations that are already installed near coverage holes presented here, working to make those valuable observations available to AMPS should be made a priority. Overall, the pressure correlation length scales are much longer than for temperature, and the current network provides adequate coverage for pressure on daily time scales. In contrast, temperature measurements leave large regions essentially unmonitored on the daily time scale, with areas of poor coverage even for weekly time averages.
In terms of seasonal and regional differences, the austral summer carries much of the correlation signal, especially in the interior. King et al. (2003) and Chapman and Walsh (2007) also found longer temperature correlation length scales in austral summer (defined as DJF) when compared to winter (JJA). However, there are local differences that vary with location and season, so that the influence of each station is not isotropic. Coastal stations explain a larger fraction of temperature variance during winter, which we hypothesize is related to persistent katabatic winds producing correlations over long distances. In contrast, interior stations have a larger signal in summer, which we hypothesize is due to relatively weaker inversion strength; with greater vertical mixing, local surface-based influence diminishes and correlation length scales are longer.
These differences illustrate the need for objective optimal network design (e.g., Huntley and Hakim 2010; Mauger et al. 2013), which depends on monitoring goals, region of interest, and the time of year. For instance, to optimize a network to measure West Antarctic weather, our results suggest that Byrd, Antarctica (80°S, 120°W), may not be the ideal location for a station due to the relatively short correlation length scales on daily time scales (Fig. 5). On the other hand, for safe aircraft operations at Byrd itself, the station is invaluable. Additionally, on monthly time scales, or for monitoring climate, the station is well correlated with the rest of West Antarctica, as also illustrated by Bromwich et al. (2013a).
In general, West Antarctica and the peninsula have relatively short CLSs on the daily time scale (Fig. 5), suggesting that more stations are needed per unit area to adequately sample temperature in these locations. West Antarctica is particularly noteworthy for being poorly sampled (Fig. 7) and an area experiencing climate-driven glaciological changes (e.g., Steig et al. 2012). The gaps identified here in the current network coverage (Figs. 6 and 7), the expense of installing and maintaining weather stations in the harsh Antarctic climate, and the relative “youth” of the network, all make Antarctica an ideal location to utilize objective optimal network design techniques. These methods can be used for different monitoring objectives, such as identifying the best station locations in West Antarctica to constrain uncertainty in climate monitoring, or to minimize forecast error near McMurdo station, for example, to aid in better forecasts for research flights and rescues.
It is important to clarify that the existing stations in Antarctica were put in place for a variety of different purposes, with both research and operational goals in mind. Although an objective, statistical analysis was not performed to determine station placement, the station locations were determined using informed decisions that took into account several logistical considerations and constraints. Some of these challenges include multiple countries with differing funding guidelines and priorities working over the continent, evolving research objectives for individual programs, safety concerns for field operations, and other practical and political motivators. The follow-on research into objectively determined station locations can provide guidance for augmenting the existing network based on specific monitoring goals; existing stations already serve a purpose that may differ from the optimized objective.
In closing, a sample of what an objective network analysis (Huntley and Hakim 2010; Mauger et al. 2013) can provide is given in Fig. 8, which shows the optimal locations for monitoring the climate of three regions of the continent that were the focus of this study, assuming no existing stations. These locations were determined by correlating the spatial-mean daily temperature of each region—the monitoring goal in this example—with every point on the grid within the three regions, separately. A Monte Carlo technique using 10 000 iterations was used for the calculation in each region, and these iterations used 100-member ensemble samples randomly drawn from the AMPS data. Figure 8 shows the percentage of samples each grid point was drawn in the 10 000 iterations, revealing “hotspots” for the ideal location of the first station in each region. For West Antarctica, the ideal location (blue to red pixels) is found in a band between Byrd and the Ross Ice Shelf. For East Antarctica, the ideal location is near the Mega Dunes area, which corresponds with both a current hole in the network to the east of Vostok (Figs. 6 and 7) and an area that has relatively long correlation length scales (Fig. 5). For the peninsula, the ideal location is located on either side of the mountains near Palmer Land. These examples highlight the likely area of the one location best suited to monitoring daily temperature within each region. While the first location can be found easily using correlations, subsequent stations are conditional on these, since the goal is to select locations that best explain the residual variance not explained by previous stations. The results of this exercise can be unintuitive, illustrating the importance of an objective approach, the results of which will be reported in an upcoming publication.
The percentage of iterations a grid point was chosen as the first station in (top) the Antarctic Peninsula, (middle) West Antarctica, and (bottom) East Antarctica using an optimal network design technique for measuring the mean daily temperature of each region. 10 000 Monte Carlo iterations were used for each region. The percentages are smaller for larger regions because the calculations draw from a larger number of grid cells. The green dots represent all of the stations available to be assimilated by AMPS, regardless of how frequently they report.
Citation: Monthly Weather Review 142, 10; 10.1175/MWR-D-13-00401.1
Acknowledgments
This study was supported by NSF Grant 1043090, awarded to the University of Washington. Jordan Powers and Kevin Manning (NCAR) provided access to both the observations and the AMPS gridded data, and for that we are grateful. The authors thank two anonymous reviewers and David Bromwich for thoughtful and useful review comments that helped strengthen this manuscript.
REFERENCES
Bromwich, D. H., J. J. Cassano, T. Klein, G. Heinemann, K. M. Hines, K. Steffen, and J. E. Box, 2001: Mesoscale modeling of katabatic winds over Greenland with the Polar MM5. Mon. Wea. Rev., 129, 2290–2309, doi:10.1175/1520-0493(2001)129<2290:MMOKWO>2.0.CO;2.
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, doi:10.1175/MWR-2881.1.
Bromwich, D. H., J. P. Nicolas, A. J. Monaghan, M. A. Lazzara, L. M. Keller, G. A. Weidner, and A. B. Wilson, 2013a: Central west Antarctica among the most rapidly warming regions on Earth. Nat. Geosci., 6, 139–145, doi:10.1038/ngeo1671.
Bromwich, D. H., F. O. Otieno, K. M. Hines, K. W. Manning, and E. Shilo, 2013b: Comprehensive evaluation of polar weather research and forecasting model performance in the Antarctic. J. Geophys. Res. Atmos., 118, 274–292, doi:10.1029/2012JD018139.
Chapman, W. L., and J. E. Walsh, 2007: A synthesis of Antarctic temperatures. J. Climate, 20, 4096–4117, doi:10.1175/JCLI4236.1.
Ding, Q., and E. J. Steig, 2013: Temperature change on the Antarctic Peninsula linked to the tropical Pacific. J. Climate, 26, 7570–7585, doi:10.1175/JCLI-D-12-00729.1.
Guo, Z., D. H. Bromwich, and J. J. Cassano, 2003: Evaluation of Polar MM5 simulations of Antarctic atmospheric of Antarctic atmospheric circulation. Mon. Wea. Rev., 131, 384–411, doi:10.1175/1520-0493(2003)131<0384:EOPMSO>2.0.CO;2.
Huntley, H., and G. Hakim, 2010: Assimilation of time-averaged observations in a quasi-geostrophic atmospheric jet model. Climate Dyn., 35, 995–1009, doi:10.1007/s00382-009-0714-5.
Jenkins, A., P. Dutrieux, S. S. Jacobs, S. D. McPhail, J. R. Perrett, A. T. Webb, and D. White, 2010: Observations beneath Pine Island Glacier in West Antarctica and implications for it retreat. Nat. Geosci., 3, 468–472, doi:10.1038/ngeo890.
King, J. C., and J. C. Comiso, 2003: The spatial coherence of interannual temperature variations in the Antarctic Peninsula. Geophys. Res. Lett., 30, 1040, doi:10.1029/2002GL015580.
King, J. C., N. P. M. van Lipzig, W. M. Connolley, and J. C. Comiso, 2003: Are temperature variations at Antarctic ice core sites representative of broad-scale climate variations? Seventh Conf. on Polar Meteorology and Oceanography and Joint Symp. on High-Latitude Climate Variations, Hyannis, MA, Amer. Meteor. Soc., 1.8. [Available online at https://ams.confex.com/ams/7POLAR/techprogram/paper_61095.htm.]
Kunkel, K. E., D. R. Easterling, K. Hubbard, K. Redmond, K. Andsager, M. C. Kruk, and M. L. Spinar, 2005: Quality control of pre-1948 cooperative observer network data. J. Atmos. Oceanic Technol., 22, 1691–1705, doi:10.1175/JTECH1816.1.
Lazzara, M. A., G. A. Weidner, L. M. Keller, J. E. Thom, and J. J. Cassano, 2012: Antarctica automatic weather station program: 30 years of polar observations. Bull. Amer. Meteor. Soc., 93,1519–1537, doi:10.1175/BAMS-D-11-00015.1.
Mauger, G. S., K. A. Bumbaco, G. J. Hakim, and P. W. Mote, 2013: Optimal design of a climatological network: Beyond practical considerations. Geosci. Instrum. Method. Data Syst., 2, 199–212, doi:10.5194/gi-2-199-2013.
Monaghan, A. J., D. H. Bromwich, J. G. Powers, and K. W. Manning, 2005: The climate of the McMurdo, Antarctica, region as represented by one year of forecasts from the Antarctic Mesoscale Prediction System. J. Climate, 18, 1174–1189, doi:10.1175/JCLI3336.1.
Nicolas, J. P., and D. H. Bromwich, 2011: Climate of West Antarctica and influence of marine air intrusions. J. Climate, 24, 49–67, doi:10.1175/2010JCLI3522.1.
Nigro, M. A., J. J. Cassano, M. A. Lazzara, and L. M. Keller, 2012: Case study of a barrier wind corner jet off the coast of Prince Olav Mountains, Antarctica. Mon. Wea. Rev., 140, 2044–2063, doi:10.1175/MWR-D-11-00261.1.
Parish, T. R., and D. H. Bromwich, 1997: On the forcing of seasonal changes in surface pressure over Antarctica. J. Geophys. Res., 102, 13 785–13 792, doi:10.1029/96JD02959.
Parish, T. R., and D. H. Bromwich, 2007: Reexamination of near-surface airflow over the Antarctic Continent and implications on atmospheric circulations at high southern latitudes. Mon. Wea. Rev., 135, 1961–1973, doi:10.1175/MWR3374.1.
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, doi:10.1175/BAMS-D-11-00186.1.
Seefeldt, M. W., and J. J. Cassano, 2012: A description of the Ross Ice Shelf air stream (RAS) through the use of self-organizing maps (SOMs). J. Geophys. Res., 117, D09112, doi:10.1029/2011JD016857.
Skamarock, W. C., and Coauthors, 2008: A description of the Advanced Research WRF version 3. NCAR Tech. Note NCAR/TN-475+STR, 113 pp. [Available online at http://www.mmm.ucar.edu/wrf/users/docs/arw_v3_bw.pdf.]
Steig, E. J., Q. Ding, D. S. Battisti, and A. Jenkins, 2012: Tropical forcing of circumpolar deep water inflow and outlet glacier thinning in the Amundsen Sea Embayment, West Antarctica. Ann. Glaciol., 53, 19–28, doi:10.3189/2012AoG60A110.
Steinhoff, D. F., D. H. Bromwich, M. Lambertson, S. L. Knuth, and M. A. Lazzara, 2008: A dynamical investigation of the May 2004 McMurdo Antarctica severe wind event using AMPS. Mon. Wea. Rev., 136, 7–26, doi:10.1175/2007MWR1999.1.
Thoma, M., A. Jenkins, D. Holland, and S. Jacobs, 2008: Modelling circumpolar deep water intrusions on the Amundsen Sea continental shelf, Antarctica. Geophys. Res. Lett., 35, L18602, doi:10.1029/2008GL034939.
Thompson, D. W., S. Solomon, P. J. Kushner, M. E. England, K. M. Grise, and D. J. Karoly, 2011: Signatures of the Antarctic ozone hole in Southern Hemisphere surface climate change. Nat. Geosci., 4, 741–749, doi:10.1038/ngeo1296.
Warren, S. G., and M. S. Town, 2011: Antarctica. Encyclopedia of Climate and Weather, 2nd ed. S. H. Schneider, T. L. Root, and M. D. Mastrandrea, Eds., Oxford University Press, 63–71.
Wilks, D. S., 2011: Exploratory techniques for paired data. Statistical Methods in the Atmospheric Science, 3rd ed. R. Dmowska, D. Hartmann, and H. T. Rossby, Eds., Elsevier, 49–60.
Note that the color scale ranges are different for each of the time averages to highlight the structure in the patterns. The lowest amount of variance explained for the monthly averages, for example, is 70%.
