CHANGES IN ANTARCTIC PENINSULA TROPOSPHERIC TEMPERATURES FROM 1956 TO 1999: A SYNTHESIS OF OBSERVATIONS AND REANALYSIS DATA

Transcription

1 INTERNATIONAL JOURNAL OF CLIMATOLOGY Int. J. Climatol. 22: (2002) Published online in Wiley InterScience ( DOI: /joc.758 CHANGES IN ANTARCTIC PENINSULA TROPOSPHERIC TEMPERATURES FROM 1956 TO 1999: A SYNTHESIS OF OBSERVATIONS AND REANALYSIS DATA GARETH J. MARSHALL, a, * VICTOR LAGUN b,1 and THOMAS A. LACHLAN-COPE a a British Antarctic Survey, Natural Environment Research Council, High Cross, Madingley Road, Cambridge CB3 0ET, UK b Arctic and Antarctic Research Institute, Russian Federal Service for Hydrometeorology and Environmental Monitoring, 38 Bering Street, St Petersburg, Russia Received 22 July 2001 Revised 27 November 2001 Accepted 27 November 2001 ABSTRACT The surface warming at Faraday station in the western Antarctic Peninsula is one of the largest observed anywhere over the last 50 years, yet the physical mechanisms driving this climate change are poorly understood. In this paper we synthesize radiosonde temperature observations from three Peninsula stations and NCEP NCAR reanalysis data in order to examine contemporaneous regional tropospheric temperature trends ( ), which may in turn help us to understand better the causes of the surface warming. The reanalysis data are utilized in two ways: (i) to provide long-term mean monthly offsets between Faraday, which ceased radiosonde observations in 1982, and two other stations in the region having more recent data, Bellingshausen and Marambio, in order to create post-1982 simulated Faraday data; (ii) after having any spurious trends and bias removed, to provide directly a monthly value for Faraday when no equivalent value from regional observations is available. Using available months of overlap, a comparison between temperature observations and simulated data suggests that the latter are a reasonable facsimile of the former. The synthesized time-series of tropospheric temperatures reveal a statistically significant mean annual tropospheric ( hpa) warming above Faraday between 1956 and 1999 of ± C year 1. Winter and summer both show a warming trend, with significance varying with height and season. Annually, the mean tropospheric warming is half that at the surface, Unlike the surface warming, the calculated tropospheric warming trend is no greater than observed at other Antarctic stations, and indeed is not significantly greater than the background global warming trend for most of the period examined. Thus, we cannot dismiss the possibility that the Peninsula surface warming may simply be a response to a global warming magnified by the observed strong regional feedback between sea-ice extent and surface temperature during winter. Copyright 2002 Royal Meteorological Society. KEY WORDS: Antarctic Peninsula; climate change; surface temperature; tropospheric temperature; reanalysis 1. INTRODUCTION Routine meteorological surface observations from the Antarctic Peninsula began in the mid-1940s. The most reliable long-term ( 50 years) temperature record from this region has been obtained from Faraday station on the western side of the Peninsula (61 15 S, W) (note that in 1996 the station became Vernadsky, when its operation passed to the Ukraine; however, as this paper utilizes the upper-air data from this station, which ceased in 1982, it will be referred to as Faraday throughout). Reliable, consistent data from Faraday began in April 1950; the annual surface temperature trend for the 50 years from is ± C year 1 (see Section 5 for the methodology for determining the 95% confidence * Correspondence to: G. J. Marshall, British Antarctic Survey, Natural Environment Research Council, High Cross, Madingley Road, Cambridge, CB3 0ET, UK; gjma@pcmail.nbs.ac.uk 1 lagun@aari.nw.ru Copyright 2002 Royal Meteorological Society

2 292 G. J. MARSHALL, V. LAGUN AND T. A. LACHLAN-COPE intervals), which is significant at the 5% level. Other regional records show that the warming is not simply confined to the western Antarctic Peninsula. Stations in the northeastern Peninsula have shown smaller but, nevertheless, statistically significant temperature increases (Skvarca et al., 1998). The region of significant warming extends northward into South America, reaching as far as 50 S (Hoffmann et al., 1997), and limited data from the southern Peninsula also point to evidence of a recent warming (Harangozo et al., 1997; Shuman and Stearns, 2001). Hansen et al. (1999) demonstrated that from 1950 to 1998 the Peninsula region underwent the greatest warming anywhere in the Southern Hemisphere, which was of a similar magnitude to the largest coincident global temperature rises, observed in regions of the Arctic (see their Plate 4a). Seasonally, the greatest 50 year warming at Faraday has occurred during the austral winter (June August: JJA) of ± C year 1 ( ), which has been linked to a decrease in winter sea-ice extent in the Bellingshausen Sea west of the Peninsula (Jacobs and Comiso, 1997; King and Harangozo, 1998; van den Broeke, 2000b). However, it is the smaller summer (December February: DJF) temperature increase of ± C year 1 that is principally responsible for the disintegration of many of the ice shelves fringing the northern Peninsula (Vaughan and Doake, 1996). There have also been changes in the regional marine ecosystem that are coincident with the warming. Peninsula populations of penguins that prefer a winter habitat within the pack ice have declined, whereas those that prefer open water have shown a marked increase and southward expansion of their range (Fraser et al., 1992; Smith et al., 1999). Despite the magnitude of the Peninsula warming, the physical processes driving it remain unclear (Vaughan et al., 2001). King and Harangozo (1998) postulated that regional atmospheric circulation changes are the most likely cause. Using mean sea-level pressure (MSLP) charts of various origins (and questionable accuracy) they showed a decrease in absolute pressures in the eastern Bellingshausen Sea during winter in the 1970s and 1980s compared with the late 1950s and early 1960s, with a resultant increase in frequency and/or intensification of northerly winds and hence warm advection over the Peninsula itself. Turner et al. (1997) revealed a statistically significant increase in the number of reports of winter precipitation at Faraday, which is mostly synoptic in origin (Turner et al., 1995), coincident with the warming. Again, this suggests a long-term increase in winter cyclonic circulation upstream of the Peninsula. However, this is not borne out by the study of Simmonds and Keay (2000), who detected a decline in Antarctic cyclone numbers west of the Peninsula, one of the few coastal regions of the continent not to show an increase. It should be remembered, though, that the Simmonds and Keay (2000) data are derived from the National Centers for Environmental Prediction (NCEP) National Center for Atmospheric Research (NCAR) reanalysis (hereinafter written as NNR) MSLP fields, which have marked spurious negative trends at southern high latitudes (Hines et al., 2000; Marshall and Harangozo, 2000), and this may have impacted upon their results. Regional oceanographic changes may also have had a role to play in the warming. Upwelling circumpolar deep water (CDW), which rises in response to the Ekman transport of surface waters (Hofmann et al., 1996), is believed periodically to flood the bottom waters (between m and the 500 m shelf bottom) of the continental shelf west of the Peninsula. Potter and Paren (1985) believed that ice melting processes taking place above the inner shelf are the driving mechanism for across-shelf CDW transport; the outflow of more buoyant surface waters produced by ice melt is replaced by onshore CDW transport at depth and the upwelled CDW, a relatively warm water mass, provides a heat source to continue the melting and hence the resultant circulation. Temperature changes in the continental shelf bottom water of 1 2 C have been inferred in the late Holocene (Leventer et al., 1996), suggesting that a recent increase in the presence of CDW on the shelf would have the potential to impact significantly on the formation of sea ice. Alterations in the general atmospheric circulation at high southern latitudes will drive the regional changes described above. The semi-annual oscillation (SAO) is a major component of climate variability in this region and is caused by the dynamically coupled but differing temperature cycles of the Antarctic continent (coreless winter) and Southern Ocean sea-surface temperatures (SSTs) at 50 S see van Loon (1967) and Meehl (1991) for more details such that the circumpolar trough (CPT) of low pressure that rings the Antarctic is both deepest and furthest south during the equinoctial seasons. In the mid-1970s the SAO began to weaken across the Southern Hemisphere (Hurrell and van Loon, 1994), including the Peninsula region, where its strength declined until the mid-1980s (van den Broeke, 1998). Van den Broeke (2000b) suggested that the

3 ANTARCTIC PENINSULA TROPOSPHERIC TEMPERATURE CHANGES 293 amplitude of the SAO could be coupled with winter sea-ice extent in the Bellingshausen Sea, which is highly correlated to Peninsula temperatures in winter months (King and Harangozo, 1998). He argued that in (recent) years, when the SAO is weaker, the northwestward migration of the CPT from April July is suppressed, causing regional negative MSLP anomalies and resultant increased northerlies a finding that agrees with King and Harangozo (1998) that limit the northward migration of sea-ice extent in winter. Van den Broeke (2000a) also found an increase in regional cloudiness in years with a weaker SAO; cloud will both reflect incident short-wave solar radiation at the top of the atmosphere (cooling) and increase the downward component of outgoing long-wave radiation (warming), but, over a high-albedo surface, such as sea ice, the second effect is significantly greater, leading to a net warming, again exerting a check on sea-ice growth (e.g. Makshtas et al., 1999). The impact of a changing SAO will be affected by contemporaneous alterations in the spatial pattern of high-latitude variability. Connolley (in press) has demonstrated that this switches between two principal modes: the wave-number 2 structure of the Antarctic circumpolar wave (ACW) (White and Peterson, 1996) and a wave-number 3 pattern with a standing rather than precessional oscillation. It seems likely that any changes in the Peninsula region, whether atmospheric or oceanographic, are influenced by the El Niño southern oscillation (ENSO). The mechanism through which ENSO impacts the Peninsula region appears to be through atmospheric Rossby wave propagation from the tropical Pacific Ocean. Experiments with atmospheric models indicate that such waves form in response to anomalous tropical convection, (e.g. Hoskins and Karoly, 1981). This occurs when SSTs exceed a certain threshold, but the sensitivity of the relationship between SSTs and tropical deep convection, and therefore Rossby wave generation, is highly dependent on large-scale vertical motions (Lau et al., 1997) and thus results in an imperfect link. Nevertheless, Trenberth and Caron (2000) demonstrated that a (negative) correlation of magnitude greater than 0.4 exists between the southern oscillation index (SOI) and MSLP in the Bellingshausen Sea, as derived from NNR data for ; when there is enhanced tropical convection (negative SOI) then MSLP is higher northwest of the Peninsula and vice versa. This correlation is significant at the 5% level, and, interestingly, on an annual basis, the change in MSLP in the Bellingshausen Sea corresponding to one standard deviation departure in the SOI is the largest observed anywhere on Earth. Further evidence for the strong relationship between Bellingshausen Sea MSLP and ENSO is provided by the study of Renwick and Revell (1999). They showed that blocking in the Bellingshausen Sea region is strongly modulated by ENSO and, using a series of barotropic modelling experiments, confirmed the link between tropical deep convection, Rossby wave propagation and regional pressure variability west of the Peninsula. Marshall and King (1998) demonstrated a clear relationship between winter 500 hpa heights in the Bellingshausen Sea and temperatures at Faraday; warm winters are associated with negative height anomalies and vice versa. These authors also noted the inexact correspondence between ENSO and the southern high-latitude signal. Note that, given the relationships described above, if a perfect association existed it would be difficult to reconcile the Peninsula warming of the last 50 years with the contemporaneous increasing frequency of the El Niño phase of ENSO as it is currently defined. However, when Harangozo (2000) examined ENSO teleconnections to the tropics using a measure of the SST change during austral winter rather than absolute values, he found that such a SST parameter was well correlated with westerlies in the central Pacific (40 55 S) and meridional flows in the Peninsula region. The latter, together with sea-ice extent at the beginning of winter, was the main factor governing sea-ice extent in the Bellingshausen Sea and temperatures on the western side of the Peninsula. In this paper we attempt to improve our understanding of the physical mechanisms behind the Peninsula surface warming by examining changes in tropospheric temperatures above Faraday coincident with the surface warming. We are interested in whether a significant warming is seen above the surface, and, if so, how its magnitude varies annually and throughout the year compared with the surface temperature rise. The 44 year temperature record from 1956 to 1999 is analysed at five pressure levels: one actually above the tropopause, in the lower stratosphere (100 hpa), and four in the troposphere (300, 500, 700 and 850 hpa). Note that occasionally the 300 hpa level may be in the lowermost stratosphere, this being most likely in winter when the tropopause is indistinct as the stratosphere cools rapidly; however, observations indicate that the 300 hpa level typically lies within the troposphere, and it will be defined as such in this paper. As the Faraday upper-air programme ended in 1982, a synthesis of observations and simulated data are necessarily employed to produce the full 44 year dataset. The latter are based on a combination of NNR and observations

4 294 G. J. MARSHALL, V. LAGUN AND T. A. LACHLAN-COPE Figure 1. Station locations of radiosonde data utilized in this study from other Antarctic stations operating upper-air programmes in the northern Peninsula region; Bellingshausen (62 12 S, W) and Marambio (64 14 S, W), are utilized (see Figure 1 for the station locations). In Section 2 of this paper the characteristics of both types of data used to produce the synthesized timeseries are depicted: the radiosonde observations and, in particular, the quality control methodology used in this study are described, together with an outline of the NNR data. The methodology employed in generating simulated Faraday upper-air temperatures from the different data sources and types is described in Section 3. Furthermore, the accuracy of the simulated Faraday data is considered by comparing them with actual observations when possible. In Section 4 the resultant synthesized Faraday upper-air temperature series is depicted in terms of its basic statistics and long-term trend on an annual and seasonal basis. Finally, in Section 5, the new upper-air temperature trends are compared with the dramatic warming observed at the surface and the findings are discussed in terms of what they reveal about the possible physical mechanisms driving the latter. 2. DATA 2.1. Radiosonde data Daily data from three radiosonde stations are utilized: Faraday, Bellingshausen and Marambio. A complete set of all available flights from Faraday was available at the British Antarctic Survey, encompassing the period A large amount of the Bellingshausen data was available from the so-called Parish dataset ; this covers all Antarctic stations from 1980 to 1993 although the data from a given station are sometimes far from complete and can be obtained from the Antarctic Meteorological Research Center (AMRC) at the University of Wisconsin. TEMP messages from the Global Telecommunication System (GTS) were available to the authors from 1988, via the UK Met Office. Further data were obtained from the archives at the Arctic and Antarctic Research Institute (AARI) to supplement/correct the other data. Unfortunately, the digital archives at the AARI do not begin until 1979; therefore, relatively few observations are available from this station before this time (although this situation is being addressed). In addition, there are gaps in the data series,

5 ANTARCTIC PENINSULA TROPOSPHERIC TEMPERATURE CHANGES 295 the largest being between December 1996 and November Marambio data have also been obtained via the Parish dataset and the GTS. Unfortunately, for the current application, relatively few radiosonde ascents from Marambio appear to have reached the GTS, meaning that there are sometimes insufficient data to produce a reliable monthly mean (see below); therefore, large gaps exist in this data series too Quality control methodology. Many studies of radiosonde data have assumed the dataset to be temporally homogeneous. However, there are multifarious reasons why such an assumption may be invalid, such as changes in any part of the equipment used to obtain upper-air measurements from radiosondes see Gaffen (1994), Parker and Cox (1995) and Gaffen et al. (2000) for details. Adjustments for these systematic errors (bias) can be achieved through comparisons with neighbouring stations (so-called buddy-checking ), operational model analyses (if these can be trusted; see Marshall (2002)), international radiosonde intercomparisons, and by modelling the thermodynamics of the radiosonde instruments themselves (e.g. Luers and Eskridge, 1998). In addition, data derived from the TEMP messages via the GTS (e.g. some of the Bellingshausen and Marambio data) may be corrupted by occasional random errors during the transmission process: incorrect coding procedures can be employed at the originating station or data may be miskeyed. These gross errors may be very large indeed. Thus it is essential that all radiosonde observations be quality controlled prior to further data analysis. The quality control process utilized in this paper employs techniques from the fields of resistant, robust and non-parametric statistics as described by Lanzante (1996). For analysis of radiosonde time-series, these have the important advantage over parametric techniques of being less affected by any outliers that may exist as a result of the random errors described above. A summary of the procedures is given below (see Lanzante (1996, 1998) for a more detailed account of the statistical theory and the explicit equations employed) Removal of random errors the biweight mean and standard deviation: For Faraday and Bellingshausen data this element of the quality control was undertaken by segregating the radiosonde timeseries of each station into individual calendar months to account for the annual cycle in temperature. For each month the biweight mean and standard deviation were estimated. Biweight estimates are weighted averages with observations having a non-linearly decreasing weight away from the distribution centre; beyond a certain distance from the centre the extreme outliers can be censored (weighting is zero). In this study that distance was set to five standard deviations from the mean. Preliminary estimates of the median and median absolute deviation were then used to determine the weights given to each observation. Once the weights of the outliers were adjusted to reduce, or indeed remove, their influence on the statistics the biweight mean and standard deviation were estimated. Finally, using these two parameters, all data further than three standard deviations from the mean were discarded. The percentages of observations removed in each station pressure-level combination are given in Table I; the fraction removed varied between 0.1 and 1.3%. Unfortunately, there are not sufficient data available from Marambio in most months to carry out this method of quality control; therefore, the raw data were plotted and subjective decisions made on whether to eliminate outliers or not. Note that the change-point test of Lanzante (1996) was carried out on the Faraday data, but as it did not result in any changes it is not described here. Table I. Percentage of observations removed as outliers during the quality control process for each station pressure-level combination Station Observations removed (%) 100 hpa 300 hpa 500 hpa 700 hpa 850 hpa Faraday a Bellingshausen a It is believed that the Faraday 500 hpa temperatures had already had some preliminary quality control undertaken.

6 296 G. J. MARSHALL, V. LAGUN AND T. A. LACHLAN-COPE Production of the monthly mean value: The last component of the quality control was to produce a monthly mean figure. Following the procedures laid down by the Meteorological Office (1979), a mean monthly value was calculated providing that at least ten data points existed for that month and that any gap between observations was less than 5 days. The numbers of monthly values produced for each given station pressure-level combination are given in Table II and the resultant time-series are shown in Figure 2. As a final assessment, buddy-checking was carried out between Bellingshausen and Marambio data when possible; the results are shown in Figure 3. It can be seen that there are two periods (late 1993 and most of 1995) when the difference between the two stations ( t; Bellingshausen Marambio) is much higher than average at both 300 and 500 hpa. At 100 hpa there is some evidence that during the second of these periods t is higher than average, although not as great as a single month in There is no evidence of any marked increase in t at lower levels. Analysis of the two station records reveals that it is the Bellingshausen temperatures that are more divergent from the monthly mean during these periods, and they also diverge from the NNR data. A check on some of the individual radiosonde ascents showed no obvious problems with the data, and the flights appeared internally consistent. Nevertheless, Bellingshausen data for these periods were blacklisted, although, for reasons outlined in Section 3.1, Marambio data are used in preference to Bellingshausen data when both exist. Note that Figure 2(b) and (c) indicate that there are other months when the Bellingshausen temperatures are higher (by 2 C) than average, such as at the end of 1997, but there are no Marambio data with which to make a direct comparison. However, these Bellingshausen observations deviate far less from the NNR data than during the two periods that have been masked, and they are included as appropriate Reanalysis data In a reanalysis, a consistent state-of-the-art atmospheric forecast model and data assimilation scheme are run on an historical archive of meteorological observations, thus removing spurious climate change caused by updates to the model code. However, variations in the amount and type of data may still cause significant jumps in the reanalysis climate, especially in data-sparse regions like Antarctica (e.g. Marshall, 2002). The NNR project is described in detail by Kalnay et al. (1996) with an overall analysis by Kistler et al. (2001). In this section only the main features of the reanalysis model and those that impact specifically on the upper-air temperature fields are outlined. The reanalysis model is based on the NCEP operational model of 10 January 1995 with a reduced horizontal resolution of T62 ( 210 km) and 28 vertical levels. As the reanalysis uses data types additional to the operational analysis a more complex quality control system was introduced. The principal resource for upper-air data in the NNR was a global radiosonde database, provided by NCAR, which was supplemented by various national archives. Temperature at the chosen standard levels is described as a class A output variable (Kalnay et al., 1996), i.e. it is influenced strongly by observed data. Other studies have demonstrated that prior to meteorological data becoming available on the GTS in 1967 many surface synoptic observations from Antarctic stations are missing from the NNR (e.g. Hines et al., 2000; Marshall and Harangozo, 2000). NCEP maintains records of the available number of upper-air observations for each month in each lat long box. The sparse distribution of Antarctic radiosonde data allows the actual number of radiosonde flights from a station to be determined from such data; those available from the three stations examined in this study are shown in Table II. Number of months for which a monthly mean temperature could be produced for each station pressure-level combination Station Number of months 100 Pa 300 Pa 500 Pa 700 Pa 850 Pa Faraday Bellingshausen Marambio

7 ANTARCTIC PENINSULA TROPOSPHERIC TEMPERATURE CHANGES 297 Figure 2. Available reliable monthly temperature observations at 100, 300, 500, 700 and 850 hpa from (a) Faraday; (b) Bellingshausen; and (c) Marambio. Note the different time scale in (b) and (c) from (a)

8 298 G. J. MARSHALL, V. LAGUN AND T. A. LACHLAN-COPE Figure 3. The difference between Bellingshausen and Marambio upper-air temperatures at 100, 300, 500, 700 and 850 hpa Figure 4. This indicates that the Faraday data are complete, with, typically, one radiosonde ascent per day. The Bellingshausen and Marambio data are far less complete and are based on available GTS data. Data from the Bellingshausen station region in the 1950s are believed to be Pilot balloon flights from the former UK station Admiralty Bay, which would not provide any temperature or height data to the reanalysis. Note that Figure 4 reveals that there are some months, for example in 1992, when there were no radiosonde observations from the Antarctic Peninsula region available to the NNR and, therefore, satellite sounder data would be the principal source of upper-air information. Satellite sounder data were first assimilated into the reanalysis in March 1975 (Jenne, 2000), but principally comprise TIROS operational vertical sounder (TOVS) data from 1979 onwards; over the Southern Ocean these were the first systematic observations available to the reanalysis and led to a major improvement in Southern Hemisphere forecasts (Kistler et al., 2001) and markedly improved the accuracy of the upper-air temperature fields above 300 hpa at Antarctic stations (Marshall, 2002). Sounder data are used to derive the vertical temperature profile, but are only employed below 100 hpa over the ocean (however, note that the position of both Bellingshausen and Marambio are defined as ocean in the model land sea mask) and only have significant weight when radiosonde data are lacking (Basist and Chelliah, 1997). An additional potential problem with the NNR concerns the Australian pseudo-observations of surface pressure (PAOBS) the product of human analysts who estimate sea-level pressure based on satellite data, conventional data, and time continuity which were assimilated with a 180 longitude error between 1979 and Owing to the sparse nature of data at high southern latitudes this had its greatest impact between 45 and 60 S, but with a decreasing influence thereafter towards the South Pole. However, it is not believed

9 ANTARCTIC PENINSULA TROPOSPHERIC TEMPERATURE CHANGES 299 Figure 4. The number of monthly upper-air ascents available to the NNR for the 2.5 lat long boxes containing the three Antarctic Peninsula stations that this will impact the present study greatly, because the significance of the errors declines rapidly from synoptic to monthly time scales (see noaa.gov/paobs/paobs 1.html). 3. METHODOLOGY 3.1. Construction of the synthesized Faraday time-series Appropriately adjusted radiosonde temperature data from Bellingshausen and Marambio stations were utilized to simulate data from Faraday, thus extending the observations beyond 1982 to produce a synthesized Faraday dataset (hereinafter known as SFD) from 1956 to Unfortunately, there is relatively little data overlap between Faraday and these stations: Marambio did not begin radiosonde flights until 1981, and although data from Bellingshausen started in 1970 there are currently not sufficient observations accessible prior to the late 1970s to permit the computation of a reliable monthly mean value. Therefore, data from the NNR have been used to produce mean monthly temperature adjustment figures between stations for each pressure level. In the few months when no radiosonde data are available, then the NNR data themselves, interpolated to the Faraday position and suitably adjusted to account for any trends and bias, are utilized. Another approach, utilizing only the NNR data to continue the Faraday observations, was also employed, but the results suggested that the former methodology was superior, and so this second approach is not described further.

10 300 G. J. MARSHALL, V. LAGUN AND T. A. LACHLAN-COPE Mean monthly temperature offsets between Faraday and both Bellingshausen and Marambio at the five pressure levels examined were derived from NNR data. Values were obtained by interpolating the monthly 2.5 lat long reanalysis data to the nearest 0.1 lat long. The offsets for Bellingshausen were calculated for ; although the upper-air data from this station extend to the early 1970s there are very few monthly means before 1979, and, moreover, Kistler et al. (2001) suggest that the introduction of TOVS data into the NNR from 1979 onwards caused sufficiently significant changes in the reanalysis for it to be considered separately from the earlier period. For Marambio, offsets were calculated from 1983 to 1999, the period over which mean monthly data were obtained from observations. The mean and root-mean-square (RMS) of the monthly temperature offsets at 500 hpa are shown in Figure 5. The mean offsets are generally smaller for Marambio and the RMS values are considerably less than at Bellingshausen at this and at all pressure levels. These findings are due to the greater synoptic activity, and hence temperature variability, on the western side of the Peninsula compared with the east. This is also seen at surface stations; e.g. the standard deviation of annual temperature at Faraday is 1.6 C, whereas at Esperanza (63 24 S, S) it is only 1.1 C. Comparison of the NNR data against monthly means derived from observations shows that for the period the reanalysis does slightly better (assuming that the observations are truth ) at Marambio than Bellingshausen at all heights except 300 hpa. These two factors, the smaller and less variable offsets and better local NNR accuracy, mean that Marambio data were used in preference to Bellingshausen data to produce simulated data in the SFD when both were available. Significant spurious trends have been shown to exist in the NNR MSLP field in the Peninsula region (Marshall and Harangozo, 2000), around the Antarctic as a whole and in the geostrophic height at upper levels (Hines et al., 2000; Marshall, 2002). It was necessary, therefore, to ensure that any false trends in upper-level temperatures of the NNR were not transferred into the SFDs. First, linear trends in the difference between Faraday observations and NNR data corresponding to that location were calculated for each month pressure level combination. This was undertaken for the entire period of available observations. Although, as mentioned previously, Kistler et al. (2001) suggest that the pre- and post-1979 periods should be considered separately in the NNR; in this case it was known that the Faraday upper-air data were available to the reanalysis during both periods (see Figure 4). The trends were then removed from the data, whether significant or not; typically 2 or 3 months at each height showed a significant trend. Most trends were positive, and very small ( 0.02 C year 1 ), indicating an increase in NNR temperatures relative to observations. This might be expected, given the negative trends in the difference between geopotential height in the NNR and observations (Hines et al., 2000; Marshall, 2002). After 1982 it was assumed that no trends existed in the reanalysis data at Faraday. Justification for this assumption was based on the insignificant trends for this Figure 5. NNR-derived monthly 500 hpa temperature offsets between Faraday and both Bellingshausen and Marambio. The horizontal lines represent the RMS temperature either side of the mean offset

11 ANTARCTIC PENINSULA TROPOSPHERIC TEMPERATURE CHANGES 301 period in the difference between observations and NNR at Bellingshausen. Moreover, after 1982 the bias in 2 m temperature at Faraday is essentially constant (Marshall and Harangozo, 2000). However, in contrast, this assumption is shown not to be valid at 100 hpa (see Section 5) because of the impact of the Antarctic ozone hole on temperatures at this level, which did not become pronounced until after Once the trends were removed the mean bias was calculated for each month and all the data then adjusted by this value. In addition to making the overall NNR bias zero (to the nearest 0.1 C ) this process also reduced the RMS errors in all but two of the 60 month pressure-level combinations. Figure 6 demonstrates how the removal of trends and bias impacted the individual monthly NNR data at 500 hpa. Thus, the SFD comprises a combination of Faraday observations until 1982 and, thereafter, a mixture of simulated Faraday data derived from Marambio and Bellingshausen data with various offsets and adjusted NNR data. The number of each of these at the different pressure levels is given in Table III. At 300, 700 and 850 hpa the proportions of the different sources within the SFD are similar; 61% for Faraday observations, 15% for Marambio-derived data, 16% for Bellingshausen-derived data and 8% directly from the reanalysis. At 500 hpa there are slightly more months when a reliable Bellingshausen monthly mean Figure 6. Comparison between raw NNR 500 hpa temperature data at Faraday and the same data after removal of trends and bias, compared with observations Table III. Number of monthly observations in SFD derived from the original Faraday observations and the three sources of simulated data Pressure level (hpa) Observations Simulated-data source Marambio Bellingshausen NNR

12 302 G. J. MARSHALL, V. LAGUN AND T. A. LACHLAN-COPE temperature value could be obtained, so the simulated data from this source increases to 17% and that derived directly from the NNR drops to 7%. However, at 100 hpa there are fewer months where a reliable mean could be derived from the Faraday and Marambio observations; the proportions of data from the SFD from these two sources are 54% and 14% respectively, with a consequent increase in the use of NNR data to 16% Testing the accuracy of the simulated Faraday data For those months between 1973 and 1982 when both Faraday and Bellingshausen observational monthly means exist it is possible to test the accuracy of the simulated SFD. Frequency histograms of the difference between the Faraday observations and simulated monthly means at the five pressure levels in 1 C binsare shown in Figure 7 and the summary statistics are presented in Table IV. Overall, the least accurate simulated data in the SFD are clearly at 100 hpa (see Figure 7(a) and Table IV), with a mean bias of 1.0 C and a value for two standard deviations ( 95% of values will be within this departure from the mean in a normal distribution) of 2.6 C. At 300 hpa the simulated SFD presents a good facsimile of observations, albeit with a slight negative bias, and the distribution of differences is approximately normal [see Figure 7(b)]. At 500 hpa there is no overall bias in the data for the 47 months examined, but Figure 7(c) reveals the non-normal distribution of differences between the simulated SFD and observations, although all lie within 2.0 C. At both 700 and 850 hpa there are slight negative biases and relatively large standard deviations in the differences, which Figures 7(d) and (e) show is principally due to the negatively skewed distribution of the differences (tail of values at the lower end). Thus, assuming that the months when the above comparison can be made are representative, then, with the exception of 500 hpa, we might expect a slight negative bias in the simulated SFD and, because the simulated data are located at the end of the SFD, with any trends derived from the latter. The above results suggest that the overall bias in the simulated data will be reasonably close to zero, the exception being at 100 hpa, but the typical standard deviation of the data is too high to produce a reasonable representation of the Faraday observations on a month-by-month basis. However, in this study we are interested in annual and seasonal trends in winter and summer. Therefore, to ascertain how well the simulated data do at these longer temporal scales they were compared with observations for years and seasons when both Faraday and Bellingshausen monthly values exist. Note that it was not possible to undertake the comparison for a complete year, so the 11 available months of 1979 were used instead. In addition three seasons winter 1979 and summer 1977 and 1978 could be used for the comparison. The bias at the five pressure levels for these four periods is given in Table V. This shows that, over the longer time periods, some of the errors in individual months are of opposite sign, leading to significantly smaller differences from the observations. With three exceptions the bias is within 1.0 C or better, and 70% of the possible pressure-level time period combinations are within 0.5 C. Note that there are no Marambio observations available for the period of overlap with Faraday observations; as the mean and RMS of temperature offsets in the NNR between these two stations is less than between Bellingshausen and Faraday (see Figure 5) we would expect that, on average, data in the simulated SFD derived from Marambio observations would be more accurate than those derived from the Bellingshausen observations. Hence the overall accuracy of the simulated SFD values is likely to be better than the above comparison suggests. Table IV. Summary statistics of the bias in SFD for months when such a comparison is possible Pressure level (hpa) n Mean ( C) 2 SD ( C)

13 ANTARCTIC PENINSULA TROPOSPHERIC TEMPERATURE CHANGES 303 Figure 7. Frequency histograms comparing SFD temperatures against observations for available months: (a) 100 hpa; (b) 300 hpa; (c) 500 hpa; (d) 700 hpa; (e) 850 hpa 4. RESULTS Linear trends were calculated from the SFD. The significance of these trends was determined using the methodology outlined by Trenberth (1984), which accounts for autocorrelation in the time-series. In this, the number of effectively independent samples n eff, rather than the number of actual samples n, isusedin the Student s t-test for significance. However, when calculating the temporal trend of a single variable, the predictor value, time, is not random, so the number of independent samples is computed using the residuals of

14 304 G. J. MARSHALL, V. LAGUN AND T. A. LACHLAN-COPE Table V. Bias in the SFD for years and summer and winter seasons when such a comparison is possible. The summer year refers to the December Period (hpa) SFD bias ( C) 100 hpa 300 hpa 500 hpa 700 hpa 850 hpa 1979 (11 months) (winter) (summer) (summer) the regression equation (e.g. Smith et al., 1996) rather than the two time-series. Confidence intervals for the trends are calculated at the 95% confidence level (see section of von Storch and Zwiers (1999)) with the degrees of freedom based on n eff. Upper-air temperature trends in the SFD at the five pressure levels studied between 1956 and 1999, together with their associated 95% confidence intervals and significance levels, are given in Table VI. In addition, data for the annual, winter and summer SFD and linear trends are plotted in Figures 8 10 respectively. Annually, the SFD has statistically significant positive temperature trends at the 5% level at all the pressure levels studied except 100 hpa, the trend of which is both negative and not statistically significant. The 100 hpa plot shows no marked change in annual temperatures until the mid-1980s (see Figure 9); a similar conclusion was reached by Randel and Wu (1999), who examined radiosonde data from eight Antarctic stations. The stratospheric cooling has been widely attributed to radiative forcing due to losses in ozone and increases in CO 2 (e.g. Ramaswamy et al., 1996). At 100 hpa both the winter and summer temperature trends in the SFD are negative (see Figures 9 and 10); the latter is about twice as large as the former and is statistically significant at the 10% level. The scenario observed in the SFD is what we might expect given the observed decrease in ozone at Faraday since the early 1980s, principally during the austral spring, but which also continues into the early summer (J. Shanklin, personal communication, 2001). The warmings observed in the troposphere ( hpa) are all of a similar magnitude, ranging from to C year 1, equivalent to a temperature rise of 1.0 to 1.4 C over the 44 years examined. Marshall (2002) examined radiosonde temperatures from four Antarctic stations and showed a gradual decrease in the warming trend with height in the troposphere at Halley and Casey stations, and although there was no clear relationship between trend and height at Mawson and Davis stations, the tropospheric trends were extremely uniform. Thus, the similar homogeneous nature of trends within observations and the SFD suggests that the latter represents a reliable facsimile of real climate change. Table VI. Annual temperature trends for the SFD and surface observations for Pressure level (hpa)/height Temperature trend ( C) Annual Winter (JJA) Summer (DJF) ± ± ± c ± a ± c ± c ± b ± ± ± b ± ± c ± b ± ± b Surface ± b ± b ± b a Significant at below 1% level. b Significant at below 5% level. c Significant at below 10% level.

15 ANTARCTIC PENINSULA TROPOSPHERIC TEMPERATURE CHANGES 305 Figure 8. Annual temperature data and linear trend at Faraday for at 100, 300, 500, 700 and 850 hpa (SFD). Note the different scale for 100 hpa In the upper troposphere, at 300 and 500 hpa, the winter and summer warming trends are similar to the annual trend, but are only statistically significant at 300 hpa. The inter-seasonal variability of both winter and summer is greater than inter-annual variability at both these heights and, therefore, is associated with larger 95% confidence limits. Note that there are two marked peaks in the 300 hpa winter data (Figure 9). The first is 1989, which remains the warmest year in the Faraday surface temperature record and the positive temperature anomaly can be observed throughout the troposphere (see Figure 9). However, the second and higher temperature in 1996 is not as marked at lower levels. Investigation shows that this point is wholly derived from Bellingshausen observations and its anomalously high value of 55.4 C results principally from the July temperature. The monthly mean was calculated from 26 daily observations and there are no obvious outliers within these. However, as mentioned in Section , a comparison of Marambio and Bellingshausen data does suggest that the latter contains errors from time to time. Assuming that July 1996 is in error, replacing it with a value from the NNR changes the winter trend to ± C year 1, significant at the 10% level. Hence a statistically significant warming remains, and, of course, we cannot be sure that the NNR value is any more reliable. At both 700 and 850 hpa the SFD winter and summer warming trends are respectively lower and higher than the annual trend (see Table VI), with only the summer warmings being statistically significant. This is due to the greater inter-annual atmospheric temperature variability nearer the surface, especially in winter, as is clearly seen in the Faraday surface temperature trends (Table VI).

16 306 G. J. MARSHALL, V. LAGUN AND T. A. LACHLAN-COPE Figure 9. As Figure 8 but for winter (JJA) data 5. DISCUSSION From the results above we can deduce that the mean annual tropospheric warming at Faraday is ± C year 1, equivalent to 1.2 C for the 44 year period from 1956 to This trend is very similar to that calculated from radiosonde observations at Halley for (Marshall, 2002). Furthermore, a mean tropospheric temperature ( hpa) trend of ± C year 1 for six Antarctic stations (not including Faraday) for was calculated by Angell (2000). These trends are all significant at the 5% level or below. Therefore, though the surface warming on the western side of the Antarctic Peninsula is considerably greater than seen elsewhere in the continent; this is not true of the contemporaneous tropospheric warming, which cannot be considered statistically different from that seen at other Antarctic stations. To put the Faraday tropospheric warming in a broader context, we can compare it to equivalent temperatures over wider areas once again derived from Angell (2000). In order to remove variability common to both the Faraday and broader-scale time-series, we assess whether the difference between the two datasets is significantly different from zero. The calculations reveal that there are warming trends in the difference between Faraday temperatures and both average Southern Hemisphere and global tropospheric temperatures for the period , both significant at the 10% level. However, further analysis shows that these trends indicating relative warming at Faraday are due principally to very low temperatures in the Angell datasets during the early 1960s. After that period the trends are close to zero and not statistically

17 ANTARCTIC PENINSULA TROPOSPHERIC TEMPERATURE CHANGES 307 Figure 10. As Figure 8 but for summer (DJF) data and for , where the year refers to the December significant, and, therefore, the troposphere above Faraday is not warming significantly faster than the global average. Table VI shows that the overall warming in the troposphere at Faraday is about half that observed at the surface. However, this ratio varies considerably between the seasons; in winter the tropospheric warming is only 20% of that at the surface, whereas in summer it is 14% greater. The inter-annual correlations between the surface warming and the SFD are shown in Table VII. As one might intuitively expect, the magnitude of the correlations decreases with height in the troposphere, and those in the stratosphere are negative. Most are statistically significant and are of a high magnitude on an annual basis and in winter. The question arises as to why the correlations in summer are much lower and whether this is related to the greater warming in the troposphere compared with the surface in this season. A possible explanation for both phenomena is that during summer the surface temperature at Faraday often exceeds 0 C the mean summer temperature for is +0.4 C and at such times the available energy is being utilized to melt sea ice and terrestrial surface snow and ice rather than simply heating the surrounding air, as it will be at higher levels. Thus, from the above, we can perhaps surmise that the high inter-annual winter correlation between Faraday surface temperatures and those in the troposphere, significant at the 1% level to 500 hpa and still at the 5% level at 300 hpa (Table VII), indicates that the same physical processes are involved in driving the variability observed at the surface and at levels throughout the troposphere. The surface warming is concentrated in winter (see Table VI), and, because of the strong relationship between winter sea-ice extent and Faraday temperatures, is believed to have been caused principally by processes leading to a reduction in the former. Further evidence for the key magnifying role of sea ice in

18 308 G. J. MARSHALL, V. LAGUN AND T. A. LACHLAN-COPE Table VII. Correlation coefficients between Faraday surface temperatures and SFD data for Pressure level (hpa) Annual Winter (JJA) Summer (DJF) b c b a a a a a a b a Significant at below 1% level. b Significant at below 5% level. c Significant at below 10% level. the surface warming is seen in the marked difference between surface and tropospheric temperature trends in this season. Thus, one hypothesis for explaining the surface warming is that an atmospheric circulation change has occurred that has led to both a reduction in sea-ice extent west of the Peninsula and a regional tropospheric temperature rise (e.g. King and Harangozo, 1998). As outlined in Section 1, van den Broeke (2000b) favoured a weaker SAO as a strong candidate to explain the surface warming; a weaker SAO means stronger northerlies and greater cloudiness, which both exert a negative influence on sea-ice growth and hence is associated with warmer Peninsula temperatures. Although this may indeed be part of the mechanism for the warming, there are a number of issues that suggest other factors are involved. For example, the SAO in the Peninsula region increased in strength between 1960 and 1975, coincident with a period of marked surface temperature increase at Faraday, the opposite relationship to that suggested by van den Broeke (2000b). Atmospheric circulation changes in the Antarctic Peninsula region may also have been caused by an alteration in the so-called Antarctic oscillation (AAO; Thompson et al., 2000). This leading mode of atmospheric variability in the Southern Hemisphere is characterized by a zonally symmetric or annular configuration, whereby geopotential height anomalies of opposite sign exist between high- and mid-latitudes. Decreases in geopotential height have been observed over Antarctica beginning in the late 1970s (e.g. Hurrell and van Loon, 1994), contemporaneous with the declining SAO. This shift to the positive mode (high index polarity) of the AAO is generally associated with cold anomalies across Antarctica, but the exception is the Peninsula region, where enhanced westerlies increase the advection of relatively warm oceanic air from the Bellingshausen Sea over the colder land (Thompson et al., 2000). However, though the observed increase in precipitation reports at the station (Turner et al., 1997) may suggest an alteration in cyclonic activity, available upper-level wind data do not reveal a circulation change at Faraday that might help to explain the surface warming. Moreover, the lack of a significant difference between the local and global tropospheric warming means that we cannot dismiss the hypothesis that the magnitude of the surface warming on the western side of the Peninsula is merely a regional magnification of a background global warming. Future work will employ diagnostic studies using data from general circulation models and the forthcoming ERA-40 reanalyses, e.g. changes in thermal advection, to examine further the climate-change processes responsible for the remarkable surface warming on the western side of the Antarctic Peninsula. ACKNOWLEDGEMENTS The authors would like to acknowledge the help of Steve Colwell for the original processing of the Faraday data. John King and John Turner kindly gave their time to read through early drafts of the manuscript. The comments of two reviewers further improved the final paper.