1. Introduction

The Southern Ocean and the Antarctic continent represent perhaps the largest spatial meteorological data voids on the globe. Climate analysis over the high southern latitudes is limited to the sparse station network, making it challenging to resolve climate signals in the high southern latitudes. Recently, various satellite datasets have become more widely distributed and employed, greatly helping to eliminate the large spatial data gaps in these regions. Satellite-derived surface temperatures (e.g., Comiso 2000), sea ice concentrations derived from passive microwave radiometers (e.g., Zwally et al. 2002), and various cloud and radiation products (including cloud track winds; Pavolonis and Key 2003; Key et al. 2003) derived from the Advanced Very High Resolution Radiometer (AVHRR) and the Moderate-Resolution Imaging Spectroradiometer (MODIS) data are just a few of the datasets utilized in recent research. However, most satellite data only extend back to 1978, forcing studies involving preceding years to largely depend on the sparse surface station network.

The National Centers for Environmental Prediction– National Center for Atmospheric Research (NCEP– NCAR) reanalysis (hereafter NCEP1; Kistler et al. 2001; Kalnay et al. 1996) project helped to solve the problem of the large data voids, particularly before the availability of the satellite data. NCEP1, unlike many of the available analyses, has the positive benefit of a fixed state-of-the-art assimilation scheme. With more observations included and the better quality control in NCEP1, these reanalysis fields were thought to be the means by which climate studies could finally be conducted over the entire southern high latitudes starting from the International Geophysical Year (IGY: 1957– 58).

Since their initial release to the public in 1996 (Kalnay et al. 1996), NCEP1 has been widely used for many climate studies in the high latitudes. Recently, NCEP1 has been utilized across Antarctica to study: the katabatic winds (Parish and Cassano 2003); trends in the circumpolar vortex/Antarctic Oscillation (Thompson and Solomon 2002; Marshall 2003; Jones and Widmann 2003); the surface energy budget (Renfrew et al. 2002; Trenberth et al. 2002); the El Niño–Southern Oscillation (ENSO) teleconnection (Bromwich et al. 2000; Genthon et al. 2003); and synoptic-scale cyclone activity (Simmonds 2000), to name only a few. The accessibility and spatial coverage of NCEP1 make it a prime choice to conduct climate studies across these regions of large observational data voids.

However, the shortage of observations still negatively affects the skill and reliability of NCEP1 in the high southern latitudes. Hines et al. (2000) observed artificial trends in the mean sea level pressure (MSLP) fields near Antarctica, due to strong positive biases that decrease with time. They identify this linear trend not only in the surface pressure, but also aloft in the 500-hPa geopotential height fields. Hines et al. (2000) demonstrated that few Antarctic surface observations were assimilated into NCEP1 prior to the availability of the Global Telecommunications System (GTS) data in 1967, despite the fact that many Antarctic stations began collecting data around the IGY. The updated NCEP–Department of Energy (DOE) Atmospheric Model Intercomparison Project-2 (AMIP-2) reanalysis (NCEP2; Kanamitsu et al. 2002), covering 1979–present (versus 1948–present of NCEP1), is very similar to NCEP1 in the period of overlap.

Marshall and Harangozo (2000) also noted the large linear trends in the MSLP fields in the Southern Ocean and West Antarctica. Their study showed that around 70°S the total decrease exceeded 12 hPa across the Southern Ocean, a trend that was statistically significant at the 1% level and was not noticeable in any nearby station observations. This trend remained largely until the 1990s, although it showed some improvement after the assimilation of the satellite sounder data beginning in the late 1970s.

Marshall (2002), in addition to the trends in the geopotential height fields, noted erroneous trends in the stratospheric temperatures in NCEP1. These trends are related to the inability of NCEP1 to capture cooling in the stratosphere associated with the seasonal ozone losses. Marshall found marked improvement in these fields after the assimilation of the satellite sounder data, which helped to further constrain the stratospheric temperatures due to the low observation density. Also noted by Marshall was a rapid drop in the East Antarctic height fields in 1993 that created a significant negative bias between NCEP1 and the observations. This sudden drop was found to be associated with the assimilation of some Australian automatic weather stations (AWS) located over the continent whose specified elevations were erroneously low.

Recently, the European Centre for Medium-Range Weather Forecasts (ECMWF) finished their 40-yr global Re-Analysis (ERA-40) spanning September 1957–August 2002 (see online at http://www.ecmwf.int/research/era/;). ERA-40 has the benefit of knowing many of the aforementioned problems encountered in NCEP1, and has taken steps to improve the skill of the reanalysis throughout the entire run. For example, ERA-40 assimilated many Antarctic stations from the start of the run that were not assimilated in NCEP1 until after the GTS data were made available (see online at http://www.ecmwf.int/research/era/Products/Archive_Plan/Archive_plan_2.html#478679 for a detailed list of the observations assimilated in ERA-40). As such, the large trends in the pressure fields (both at the surface and aloft) are expected to have been corrected in the new ERA-40 reanalysis. Further, ERA-40 is a second-generation reanalysis, having been preceded by the 15-yr Re-Analysis (ERA-15:1979–93) completed earlier by ECMWF (see Gibson et al. 1997, and references therein).

This paper compares the temporal skill in the new ERA-40 system compared to the widely used NCEP1 system. Section 2 describes the data and methods employed. Section 3 details the annual evolution of the overall (1958–2001) skill, while section 4 examines the change of skill throughout time. Section 5 extends the comparisons beyond Antarctica and the Drake Passage. Observation counts in NCEP1 are presented in section 6 to compare the temporal changes in skill with the trend in the quantity of observations assimilated into NCEP1. Section 7 provides a discussion, and conclusions are drawn in section 8.

2. Data and methodology

To validate each reanalysis, monthly mean observations from the stations in Table 1 (Fig. 1 displays the locations) were compared with the reanalysis values. Observations of MSLP were compared for all stations, whereas surface temperature was compared only for Antarctic stations (1–9 in Table 1), and 500-hPa geopotential height compared for three Antarctic stations (Casey, Halley, and McMurdo). The stations in Table 1 were chosen based on data completeness, yet some comparisons are unavoidably affected by poor data quality and are identified when needed. The station observations (both surface and upper air) for Antarctica were obtained from the British Antarctic Survey READER project Web site (http://www.antarctica.ac.uk/met/READER/;). The MSLP data for the remaining stations were obtained until 1998 through the NCAR ds570.0 dataset, with the recent years completed from data available through the National Climatic Data Center (NCDC; see online at http://www.ncdc.noaa.gov/oa/ncdc.html).

Both NCEP1 and ERA-40 are available on a 2.5° by 2.5° grid every 6 h, although both are run at higher resolutions (T-159/125 km for ERA-40 and T-62/209 km for NCEP1) and downgraded to a 2.5° resolution. ERA-40 contains 60 vertical levels (23 standard pressure levels) compared to the 28 vertical levels (17 standard pressure levels) of NCEP1. Monthly means were calculated from the reanalysis data and were bilinearly interpolated to the observational station location to within a tenth of a degree as a means of validating the performance of the two reanalyses. The use of monthly mean data allows for a basis of comparison between this study and the appraisal studies identified earlier. Very similar findings are produced using the 6-hourly data although the statistics are more influenced by the observational data completeness at this shorter time scale.

Statistics from the monthly averaged reanalysis and surface data for the MSLP field were examined to compare the evolution of overall (1958–2001) skill in ERA-40 and NCEP1. In each case, the correlation coefficient, bias, and root-mean-square error (rmse) are calculated from the station observations and the reanalysis values. Here bias refers to the mean reanalysis value over the given period minus the mean observed value. Rmse is defined as the square root of the mean-squared difference between the extracted reanalysis values and the observations, and effectively combines the errors of low correlation and high bias into one statistic.

3. Overall (1958–2001) comparison between ERA-40 and NCEP1 MSLP

Annual cycles of these statistics are presented in Fig. 2 for the high-latitude stations (mostly Antarctic) listed in Table 1 (stations 2–10), with the exception of Amundsen–Scott station, which lies well above sea level and therefore does not record MSLP. The time series for the correlations show a marked decline in ERA-40 during the austral winter (Fig. 2a). At nearly every station, the lowest correlations are observed roughly from June to August, although there are a few periods of low correlation at other times. The correlations are lower than 0.6 at four of the eight stations, namely, Casey, Mirny, Mawson, and Punta Arenas. With 13 degrees of freedom, correlations exceeding ±0.45, ±0.52, and ±0.65 are significant at the 90%, 95% and 99% levels, respectively. One might assume that the low correlation values are associated with problems in the handling of seasonal sea ice since the correlations show the lowest skill when sea ice is extensive. Deficiencies due to the handling of sea ice can arise because early sea ice data are based on the model climatology; there are few reliable sea ice observations before the modern satellite era. However, the low correlation problem extends to Punta Arenas, north of the maximum sea ice extent, suggesting that sea ice alone is not the main factor.

The correlations in NCEP1 (Fig. 2b) remain fairly high (above 0.8) throughout the year and do not show as strong a seasonal dependence for all stations excluding the East Antarctic stations. At these locations (Dumont D'Urville, Mawson, Casey, Mirny), correlations are at some points lower than those observed in ERA-40. This is particularly true in September, when the correlations at all four East Antarctic stations are at their respective lowest points. The findings are in agreement with Hines et al. (2000) who found large biases in the East Antarctic stations, largely due to the fact that the Antarctic station surface data were not assimilated into NCEP1 until the GTS data were made available in 1967. Notably, NCEP1 performs better at the other stations throughout the year, particularly in June–July–August (JJA) when ERA-40 has the most distinct problems in capturing the variability.

A slightly different outlook is presented in ERA-40 when looking at the long-term biases (Fig. 2c). Although the biases tend to reach their peak around JJA, consistent with the period of low correlation, they are small: on average less than 2 hPa too high. Relatively large deviations occur at the problem stations noted earlier, although ERA-40 reasonably captures the magnitude of the MSLP at Punta Arenas, evident from the comparably low bias at this station.

NCEP1 produces a drastically different picture (Fig. 2d). Here, at the four East Antarctic stations, the long-term biases on average are around 8 hPa too high in JJA. These four stations indicate a strong seasonal cycle in skill, as the biases are near zero during the austral summer, in agreement with previous studies (Marshall and Harangozo 2000; Hines et al. 2000). There is still a seasonal cycle of the biases at McMurdo and Halley, however the magnitude of the error at these two stations is roughly half that compared to the East Antarctic stations (∼4 hPa). For the stations in the vicinity of the Drake Passage, NCEP1 has consistently low biases near zero, and is actually performing with a higher degree of skill in this location than ERA-40.

The rmse highlights the apparent problems in both reanalyses (Figs. 2e and 2f), showing that the season of lowest skill is in the austral winter. Overall, the long-term rmse in ERA-40 is roughly half of that in NCEP1 during JJA. During DJF, the two reanalyses are comparable and closely follow the observations. Notably, the equinoctial seasons show a much stronger degradation in NCEP1 brought about by the very large long-term biases across Antarctica that peak in austral winter.

It is noteworthy that in the stations near the Drake Passage region (Orcadas and Punta Arenas) ERA-40 is generally outperformed by NCEP1. In this region, NCEP1 overall has a higher or comparable correlation, lower bias, and lower rmse. This result is somewhat surprising due to the fact that the Drake Passage region is the area with the greatest density of station observations poleward of 50°S. This would perhaps indicate that a relatively large density of station observations does not constrain ERA-40 to the extent that it controls NCEP1; this topic will be addressed further in section 7.

Altogether, Fig. 2 would favor ERA-40 over NCEP1. Although the correlations are lower during the winter in ERA-40, they are quite comparable throughout the rest of the year, and much better in September. The bias and rmse are generally lower in ERA-40. However, as these statistics are calculated using the entire years of complete overlap, 1958–2001, they can fail to capture the changes in skill with time. Particularly, nearly 50% of the 1958–2001 time interval lies in the modern satellite era (post 1978), and as such the statistics presented in Fig. 2 are weighted by the performance during this time period.

4. Time evolution

Here the statistics are presented using 5-yr windows. In this method, each parameter is calculated as before over a 5-yr span, and then edged forward a year and recalculated. This allows one to directly observe the temporal evolution in the skill as well as the impacts of assimilating satellite data. Because the austral winter was shown to be the most problematic season in terms of skill, these time series are constructed for only the JJA data, thus giving 15 individual months within a 5-yr window. The effect of assimilating satellite data is visualized by dividing the comparisons into three distinct eras of data assimilation noted in ERA-40. The first era spans 1958–72 and represents the 15 yr before any satellite data were assimilated into ERA-40. The second era, 1973–78, represents the years when the Australian surface bogus pressure data (PAOBS) were assimilated into ERA-40 from gridded Australian surface pressure analyses (A. Simmons 2004, personal communication). More importantly, during this period satellite sounder data were first assimilated into ERA-40. The Vertical Temperature Profile Radiometer (VTPR) sounding data were assimilated into ERA-40 starting on 1 January 1973, while the Television Infrared Observation Satellite (TIROS) Operational Vertical Sounder (TOVS) data entered in late 1978. The final era, 1979– 2001 represents the complete years of overlap between ERA-40 and NCEP1 when a vast array of satellite and conventional data including drifting buoys and commercial aircraft observations were available to both reanalyses, and is hereafter identified as the modern satellite era. We also observe the skill for other variables apart from MSLP.

5. ERA-40 performance in southern midlatitudes

Up to this point, ERA-40 has been shown to have some major shortcomings before the assimilation of satellite data around Antarctica where there are known deficiencies in NCEP1. Due to the strong convergence toward higher skill with the increasing assimilation of satellite data, the projected reasoning for the observed errors is most likely due to the dependence of ERA-40 on satellite observations. For this conjecture to hold true, ERA-40 would need to demonstrate the same patterns in skill across the island stations in the Southern Ocean, where large spatial gaps in data exist as in Antarctica. To verify the claim, MSLP statistics for seven southern midlatitude stations (11–17) from Table 1 are presented in Fig. 7 using the 5-yr running window method as above. Unfortunately, many of the records for these stations are incomplete, and therefore the statistics for these stations are only displayed when over 80% of the observations are included in each 5-yr window. For the most recent decades, this limits the stations to only Easter, Gough, and Marion Islands.

Overall Fig. 7 shows the same structure noted in Fig. 3. The correlations in ERA-40 (Fig. 7a) are very low in the presatellite era and are negative at some locations in the mid-1960s; however, NCEP1 also has quite low correlations (Fig. 7b) unlike those noted at the Antarctic stations. A bit surprisingly, the biases in ERA-40 (Fig. 7c) are larger and cover a greater range than in NCEP1 (Fig. 7d), a problem hinted at by Marshall (2003) when reconstructing the zonal pressure index at 40°S. The larger biases and low correlations lead to higher rmse errors in ERA-40 overall (Figs. 7e,f). Again, the trend in the skill is largely a factor of the assimilation of satellite data in ERA-40, with values converging within acceptable ranges (within the range of measurement error) almost exactly at 1979. Problems at Easter Island in 1980–84 and at Gough Island in 1997–2001 are apparent in both reanalyses and are therefore likely to be an observational data-quality issue rather than a problem with the reanalyses.

Examination of stations on or near the continental mainland or island stations with frequent ship or air traffic (i.e., Macquarie Island, Campbell Island, Chatham Island are such less isolated stations) across the Southern Hemisphere is warranted to ensure that the problems during the austral winter are not a gross deficiency in ERA-40, but are instead related to the quantity of the observational data assimilated. Due the fact that stations over the continental mainlands (Australia, New Zealand, South America) have a greater spatial density of observations than the Drake Passage region, it is expected that there will be significant improvements in the overall skill of ERA-40. Eight continental stations (18–25) were selected based on their location (farthest south, global representation) and data quality/completeness and are listed in Table 1. Statistics (using 5-yr running windows as before) are displayed in Fig. 8.

Clearly there is a large improvement in skill at these selected stations. Although the correlations of ERA-40 versus observations in Fig. 8a suggest a similar problem as in the Antarctic and Southern Ocean stations, Fig. 8a is plotted on a different scale; correlations are consistently above 0.5. Furthermore, of the four stations with the lowest correlations, two are islands south of New Zealand (i.e., Campbell Island and Macquarie Island) and are thus more influenced by the lack of nearby observations as at stations 10–16 from Table 1 (previous section). Additionally, the problems at Macquarie Island in Fig. 8a in the early stages also appear in NCEP1 (Fig. 8b), and thus could be compromised by observational data-quality issues independent of the reanalyses. However, ERA-40 still seems to have problems with the correlation at Cape Town, South Africa, and Buenos Aires, Argentina, which are not readily explained, especially since correlations at Hobart, Tasmania, and Christchurch, New Zealand—not on major continents— are consistently at or above 0.95 (Fig. 8a). The noticeable dip in correlation values in Figs. 8a,b at Cape Town right around 1979–83 is also likely due to the quality of the observational data.

The biases are small throughout and comparable. It can be argued that ERA-40 has a lower bias in the modern satellite era, with the discrepancies observed in both reanalyses in the last 5 yr or so not necessarily reflecting the true skill of the reanalyses. The trend in skill is seen in the rmse plots (Fig. 8e,f); with values in both reanalyses converging to less than 0.5 hPa, well within measurement error. Even in the initial stages of ERA-40, the rmse values are still reasonable (Fig. 8e), suggesting that it is really the lack of data that is negatively affecting ERA-40 before 1973 and not some gross error. However, NCEP1 does appear to have a slight edge in performance before the mid-1970s (Fig. 8f).

During the early years and in data-sparse regions such as Antarctica and the Southern Ocean island stations, the biases are much smaller in ERA-40 than in NCEP1, however, the correlations are generally higher in NCEP1. The small biases and simultaneous low correlations in ERA-40 demonstrate that ERA-40 relies on a more representative model climatology than NCEP1. NCEP1 captures more of the variability, but the model climatology is not representative of the observed conditions, thus creating large biases. The contrasts seen here between ERA-40 and NCEP1 are thus not only a factor of the differing assimilation schemes, but also a result of the better background fields in ERA-40 compared to NCEP1.

6. Observation counts

The NCEP–NCAR reanalysis Web site (http://wesley.wwb.noaa.gov/reanalysis.html) contains diagnostic software that allows a user to display the average number of observations assimilated into the reanalysis per 2.5° by 2.5° grid box, a product that is currently unavailable for ERA-40. Although the reanalysis projects use fixed assimilation schemes, the quantity of observed data changes significantly with time. To show the temporal and seasonal dependence of the number of observations available, plots were made over the coastal Antarctic domain (all longitudes between 60° and 80°S). Figure 9 shows the plots for austral winter (Fig. 9a) and austral summer (Fig. 9b). ERA-40 should be based on approximately the same observations as NCEP1, with the addition of the surface Antarctic stations that were absent in NCEP1 prior to the inclusion of the GTS data in 1967. It is likely that ERA-40 might also have had extra radiosonde data compared to NCEP1; additionally, concerted efforts were made to assimilate more of the early satellite data (VTPR) into ERA-40 compared to NCEP1.

As expected, there is both a strong seasonal and temporal dependence to data availability around Antarctica. During the polar winter, there were essentially no observations except the radiosonde data that were assimilated into NCEP1 before the mid-1960s (Fig. 9a). Here, ship data refer to both ship observations and buoy observations. Thus, the large spike seen in 1979 corresponds to the First Global Atmospheric Research Program (GARP) Global Experiment (FGGE) release of many drifting buoys transmitted on the GTS across the high southern latitudes. Except for this peak, the JJA ship data occur much less than radiosonde observations until the most recent decades. In the polar summer, ship observations accounted for over 80% of the total observations assimilated into NCEP1 prior to the inclusion of the surface data, after which ship observations still accounted for roughly 20%–30% of the total assimilated data (Fig. 9b). The large portion of ship data is most likely due to the continued presence of Antarctic buoy data through the World Climate Research Programme (WCRP) International Programme for Antarctic Buoys (IPAB) (WCRP IPAB 2002). Overall, it is evident that the majority of assimilated data in this region comes from the surface station data, especially in winter when surface observations account for ∼80% of the total assimilated data (Fig. 9a).

The plots in Figs. 9a,b were presented without the inclusion of the satellite data. In ERA-40, satellite-derived temperatures from the VTPR were first assimilated in January 1973, with the inclusion of TOVS in 1978; some satellite data entered NCEP1 prior to TOVS. Figure 9c displays the sum of the total observations (including the inferred surface pressure data from satellite images, known as PAOBS) assimilated into the NCEP1 for both JJA and December–January–February (DJF) plotted simultaneously for the sums with and without the satellite data. In the early 1970s, the total observation curve breaks away for each season, indicating that at least some of the VTPR data were assimilated into NCEP1. In the late 1970s the two significantly diverge, showing that in the modern satellite era over 85% of the assimilated data is information from the satellites. Such a dramatic increase in observation counts should be reflected into the reanalysis system, as many spatial gaps previously unmeasured are now filled through satellite data. As shown in Figs. 3 and 5–8, NCEP1 does not show a shift in improvement associated with the abundance of satellite data. However, ERA-40 performance reflects this change in observation density, and is well-guided and constrained by the assimilation of the satellite data during the last two decades, to the point at which it maintains a higher level of skill than NCEP1. The large distinctions in data quantity between winter and summer disappear in the modern satellite era (Fig. 9c), as the satellite inclusive series are similar in magnitude; isolated periods even exist when there are more observations in JJA than DJF. As expected, both ERA-40 and NCEP1 perform comparably well in both winter and summer over the last few decades.

Although the plots in Fig. 9 show counts produced over Antarctica for NCEP1, ERA-40 does produce global spatial plots and radiosonde time series plots for select geographical locations on the ERA-40 project Web site (http://www.ecmwf.int/research/era) under the section on monitoring. A project report series by Simmons et al. (2004) clearly shows the strong increase in observations through the modern satellite era as seen here for NCEP1. Their study also shows a decrease in observation counts in the Southern Hemisphere from 1958 to 1966. This may help to explain the decrease seen here in ERA-40 correlations during these times. Similar to this study, Simmons et al. (2004) find that the skill of the 2-m temperature analyses in ERA-40 in the Southern Hemisphere drastically increases after 1978. This increase is related not only to the inclusion of satellite data, but also to new surface observations from drifting buoys and increased data from commercial aircraft.

7. Discussion

It is important to note that the comparisons are only made at single points in the Southern Hemisphere where observational datasets are available; comparisons for the Northern Hemisphere in data-sparse areas are beyond the scope of the current study. Yet, this study still neglects the large differences occurring between NCEP1 and ERA-40 over the data-sparse Southern Ocean where large data gaps also exist. For example, average differences for all months between the 500-hPa geopotential height fields in NCEP1 and ERA-40 in the South Pacific region prior to the 1980s can be as large as ∼50 gpm (Fig. 10). However, the differences fluctuate quite drastically. By averaging the ERA-40 minus NCEP1 500-hPa geopotential height difference in a box in the South Pacific for all months (60°–70°S, 130°–150°W, i.e., the center of greatest difference in Fig. 10), a time series of bias between ERA-40 and NCEP1 is produced (Fig. 11). There are many events when the 500-hPa level in ERA-40 is over 100 gpm lower than in NCEP1 and in July 1959 the difference is >200 gpm. Without data to verify either situation projected by the reanalyses, there is no objective way to discern whether ERA-40 or NCEP1 is producing the more accurate representation. These problems persist throughout the depth of the troposphere, indicated by similar but even larger differences in the geopotential height field at 300 hPa (not shown).

Clearly, the large deficiencies noted here at the stations before the satellite era reflect even larger deficiencies in the data-void South Pacific and other oceans. This problem was alluded to in Bromwich et al. (2000), who demonstrated that over the island stations in the southern midlatitudes the differences between the ERA-15 reanalysis and the ECMWF operational analyses are generally less than half of that over the observation-sparse Southern Ocean regions, such as the South Pacific. Cullather et al. (1996), using sea level pressure observations from independent ship data in the Southern Ocean between 120°W and 180° from 1980 to 1994, demonstrated that the ECMWF operational analyses closely follow the observations in this data-sparse region. More recently, King (2003) demonstrated that the ECMWF operational analyses show very close agreement with independent surface pressure observations in the Bellingshausen Sea region for February–May 2001. Both Cullather et al. (1996) and King (2003) indicate again that the ECMWF operational analyses (that are very similar to ERA-40) are well constrained by the satellite data over otherwise data-sparse locations. However, their comparisons say nothing of the quality of these fields prior to the modern satellite era, which from Fig. 10 are significantly different than NCEP1.

8. Conclusions

The results here demonstrate significant shortcomings in both ERA-40 and NCEP1 before ∼1970. In ERA-40, very low and even negative correlations severely limit the reliability of ERA-40's surface and upper-air fields, although there is a marked improvement in the bias and rmse of these fields with time. NCEP1, however, has large biases in the MSLP fields and artificial trends in the high-latitude time series. It is shown that these problems are the largest during JJA, coincident with the small quantity of assimilated observations into both reanalyses. The problems noted here extend into the midtroposphere above the Antarctic continent, and thus are not solely a result of the reanalyses' ability to adequately resolve the Antarctic surface topography.

Although ERA-40 shows low skill in its early years, the improvements with the assimilation of satellite data are remarkable. Of all the statistics used here, ERA-40 shows impressive adjustments as the quantity of assimilated satellite data increases, converging to a skill level during the modern satellite era (post 1978) that puts its performance level above NCEP1. NCEP1, on the other hand, appears to be more constrained by the abundance of surface and radiosonde (conventional) data. Its higher performance than ERA-40 in the relatively dense data areas of the Drake Passage, along the southern extents of the major continents (Cape Town and Buenos Aires), and the southern islands indicate this fact. Differences in the assimilation schemes between ERA-40 and NCEP1 likely account for the large portion of the changes noted in this study. It is clear how the assimilation system handles the satellite data in ERA-40; with NCEP1, the inference as to what the assimilation system is doing is less obvious as little improvement occurs at the start of the modern satellite era. Also, the model climatology is clearly better in ERA-40 than in NCEP1. This is shown throughout this study by the low biases in ERA-40 in the early years, when ERA-40 is strictly following the background fields, therefore explaining the low correlations (A. Simmons 2004, personal communication). Conversely, NCEP1 had much larger biases when observations were sparse, and thus the statistics reflect a poorer model climatology.

The reliability of NCEP1 and ERA-40 before the early 1970s is questionable, however, there is no doubt that ERA-40 does an excellent job after ∼1978. Also, most of the results shown here detail the problems during the austral winter. In DJF, ERA-40 and NCEP1 perform with much higher skill, comparable to the skill seen in the more observationally dense continental stations during winter (Fig. 8). As such, austral summer studies can be extended back into the earlier years of these reanalyses (with a working knowledge of the limitations); however, care must be exercised in using the early data across the high southern latitudes and Antarctica for the other seasons, especially winter. The improvements in the assimilation scheme observed in ERA-40 by its large adjustment to the satellite data clearly indicate that reanalysis projects are taking steps in the right direction. However, additional efforts are needed in enhancing the observational (both conventional and satellite) database and in tuning the assimilation schemes before reliable data assimilation can be conducted for the presatellite era during the nonsummer months.