JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 109, D04103, doi:10.1029/2003JD004178, 2004
A 15-year West Antarctic climatology from six automatic weather
station temperature and pressure records
David B. Reusch and Richard B. Alley
Department of Geosciences and Earth and Mineral Sciences Environment Institute, Pennsylvania State University,
University Park, Pennsylvania, USA
Received 23 September 2003; revised 4 December 2003; accepted 16 December 2003; published 21 February 2004.
[1] Apart from a small number of mostly coastal stations with human observers, automatic
weather stations (AWS) are the dominant source of direct measurements of near-surface
climate parameters on the West Antarctic ice sheet. To help alleviate the shortage of surface
meteorological data in Antarctica and to better exploit this invaluable data resource, we have
used artificial neural network (ANN)-based techniques to extend and fill data gaps in
selected AWS records. Climatological analyses of the complete 15-year temperature and
pressure records (1979–1993) are reported here for six West Antarctic AWS sites (Siple
Station, Byrd, Lettau, Marilyn, Elaine, and Ferrell) spanning 90 of longitude. Three sites
(Siple, Lettau, and Marilyn) show significant warming trends during the austral summer
season (December–February). Warm anomalies of 2–5C appeared at all sites during 1980
(winter) and 1989 (winter-spring) with a cold anomaly of up to 3C during fall 1982.
Average intersite seasonal correlation comparisons are highest during fall and winter and
lowest in summer; Byrd correlates with no other sites in summer. All sites have short, sharp
temperature transitions during the fall and spring seasons and a relatively stable winter
season with occasional early to midwinter warmings. Typical winter season conditions are
present for up to 5–6 months of the year. An exception occurs at Byrd during 1988 and 1989,
when climatologically normal winter conditions did not appear to become fully established.
Mean monthly temperatures during this unusual period were 5–10C above
INDEX TERMS: 3309 Meteorology and Atmospheric Dynamics: Climatology (1620); 3344
normal.
Meteorology and Atmospheric Dynamics: Paleoclimatology; 3349 Meteorology and Atmospheric Dynamics:
Polar meteorology; 9310 Information Related to Geographic Region: Antarctica; KEYWORDS: climatology,
neural networks, polar
Citation: Reusch, D. B., and R. B. Alley (2004), A 15-year West Antarctic climatology from six automatic weather station
temperature and pressure records, J. Geophys. Res., 109, D04103, doi:10.1029/2003JD004178.
1. Introduction
[2] Direct measurements of Antarctic surface meteorology are difficult to obtain yet are essential to the larger
problem of improving interpretation of ice core-based
paleoclimate records and providing baseline data to climate
models. Unfortunately, the conditions that make an ice sheet
a good recorder of climate also make it inhospitable for
humans and their weather instruments. Thus meteorological
records from these regions are sparse and suffer greatly in
comparison to more temperate regions, particularly in terms
of continuity. An estimate of meteorological data availability gives one surface observation for every 413,000 km2 and
one upper air observation for every 1,651,000 km2 in the
Antarctic [Bromwich and Cassano, 2001]. These values are
14,000 km2 and 103,000 km2, respectively, for the United
States, or more than an order of magnitude higher spatial
resolution than in Antarctica. Automatic weather stations
(AWS) currently provide the best source of direct observaCopyright 2004 by the American Geophysical Union.
0148-0227/04/2003JD004178$09.00
tions away from the mostly coastal research stations (e.g.,
McMurdo). The United States Antarctic Program’s (USAP)
AWS network, run by the University of Wisconsin-Madison
[Stearns et al., 1993], currently includes over 50 Antarctic
sites with the highest density and longest coverage in the
Ross Island/McMurdo region and areas of the Ross Ice
Shelf and Siple Coast. The longest (and most continuous) of
these records date to 1980 and 1981 (Byrd and Ferrell,
respectively, Figure 1) but many of the other records are
substantially shorter. Furthermore, the harsh climate and
limited opportunities for repair/service (restricted by logistics to austral summer) also lead to sometimes quite extended (seasonal or longer) gaps in the records. Both of
these issues, shortness of the overall record and intermittent
data gaps, have been successfully addressed using prediction based on artificial neural networks (ANNs) [Reusch
and Alley, 2002]. This approach, inspired by climate downscaling work in temperate latitudes [e.g., Cavazos, 1999;
Crane and Hewitson, 1998], uses GCM-scale upper air data
from the European Centre for Medium-Range Weather
Forecasts (ECMWF) 15-year reanalysis (ERA-15) to predict
AWS temperature and pressure observations at 6-hourly
D04103
1 of 28
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
D04103
awsdata.html). Many of the AWS are also part of the SCAR
(Scientific Committee on Antarctic Research) READER
(Reference Antarctic Data for Environmental Research)
project (http://www.antarctica.ac.uk/met/READER/).
READER seeks to create and maintain a high-quality,
long-term data set of monthly and annual mean surface
and upper air meteorological observations from both
manned and unmanned Antarctic observing systems [Turner
et al., 2004]. Ready access to the long-term station records
available through READER provides an opportunity to put
the climatologies developed in studies such as this into a
larger spatial and temporal context. Further work on this
aspect is planned.
[5] In section 2 we describe the data used in our ANNbased prediction system. An overview of the ANN architectures and training methods used is given in section 3 (for
full details see Reusch and Alley [2002]). Section 4 presents
the synthesized temperature and pressure records for the six
AWS of this study with an analysis of predictive skill.
Section 5 provides a comparison of the ANN-based results
to ERA-15 surface data and analyses of temporal and spatial
variability.
Figure 1. Site map showing AWS sites of this study and
other locations mentioned in the text.
resolution. ANNs trained on available observations provide
predictions of missing observations due to instrument failure and during periods before and after the station was
active at the site. With the resulting complete 15-year
temperature and pressure records at six sites, it is possible
to conduct climatological studies with greater confidence in
the results.
[3] Past studies using AWS data in West Antarctica have
generally focused on the Ross Island/McMurdo region [e.g.,
Holmes et al., 2000], the western Ross Ice Shelf [e.g.,
Bromwich and Stearns, 1993] and the Siple Dome/Siple
Coast area [e.g., Das et al., 2002; Shuman and Stearns,
2001]. The six AWS sites of this study (Table 1) are located
in two distinct settings (Figure 1): the Ross Ice Shelf
(Lettau, Marilyn, Elaine and Ferrell) and the West Antarctic
Ice Sheet (Siple and Byrd). The ice shelf group is at
elevations of 50– 80 m above sea level. Locations of this
group range from fairly close to the ice shelf front (Ferrell)
to close to the Transantarctic Mountains (Marilyn and
Elaine) with Lettau at an intermediate position. The two
ice sheet sites are at substantially higher elevation (up to
1500 m a.s.l. for Byrd). Thus the six sites together have
much behavior in common but can be expected to sample
quite different aspects of the West Antarctic surface climate.
The wide spatial distribution and generally long operational
periods together provide a broad sampling of the West
Antarctic near-surface atmosphere (and a more extensive
test of the ANN-based approach).
[4] Although the six AWS of this study are widely
distributed, they represent a relatively small fraction of all
the AWS deployed in West Antarctica. More than 50 AWS
are currently active in the network managed by the University of Wisconsin-Madison since 1980 [Stearns et al.,
1993]. Another 40+ AWS no longer in operation are also
archived at their Web site http://uwamrc.ssec.wisc.edu/aws/
2. Data
[6] Two data sets were used in this work: the USAP AWS
network and the ERA-15 reanalysis product. ERA-15 provided a forecast model-based picture of the atmosphere at
moderate spatial resolution. The AWS data were both a
target for the ANN methodology, to produce complete
records during the ERA-15 period, and a source for the
ensuing climatological analysis of West Antarctic surface
climate.
2.1. AWS Data
[7] The main source of direct meteorological data in West
Antarctica is the USAP network of AWS maintained by the
University of Wisconsin-Madison since 1980 [Stearns et al.,
1993]. All stations provide near-surface air temperature,
pressure, and wind speed and direction; some stations also
report relative humidity and multiple vertical temperatures
(e.g., for vertical temperature differences). The height of the
main instrument cluster above the snow surface is nominally
3 m but this distance changes with temporal variability in the
snow surface from accumulation, settling and removal.
Additional tower sections are added over time to offset snow
accumulation. Pressure is calibrated to ±0.2 hPa with a
resolution of approximately 0.05 hPa. Temperature accuracy
is 0.25– 0.5C with lowest accuracy at 70C, i.e., accu-
Table 1. AWS Location Data, Installation Dates, and Distance to
ERA-15 Grid Points
Station
Byrd Station
Elaine
Ferrell
Lettau
Marilyn
Siple
2 of 28
a
Latitude Longitude
80.01S
83.13S
77.91S
82.52S
79.95S
75.90S
119.40W
174.17E
170.82E
174.45W
165.13E
84.00W
Elevation,
Distance,a
m
Date Installed
km
1530
60
45
55
75
1054
February 1980
January 1986
December 1980
January 1986
January 1987
January 1982b
Distance to the nearest ERA-15 grid point.
Siple AWS was removed in April 1992.
b
11.5
71
49.5
8.3
6.1
103.9
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
D04103
Table 2. Intersite Distancesa
Siple
Byrd
Lettau
Elaine
Marilyn
Ferrell
Siple
Byrd
Lettau
Elaine
Marilyn
Ferrell
–
919
1777
1877
2221
2310
919
–
929
1065
1359
1412
1777
929
–
172
445
579
1877
1065
172
–
382
583
2221
1359
445
382
–
257
2310
1412
579
583
257
–
a
Distances are given in kilometers.
racy decreases with decreasing temperature (M. Lazzara,
University of Wisconsin-Madison, personal communication, 2001). The data used here are from the 3-hourly
quality-controlled data sets available at the University of
Wisconsin-Madison Web site http://uwamrc.ssec.wisc.edu/
aws/). A 6-hourly subset of these data (for 0, 6, 12 and 18
UTC) is used to match ECMWF time steps (see below).
[8] The AWS used in this study (Figure 1) are distributed
across West Antarctica. Tables 1 – 3 provide further information on locations (Table 1), intersite distances (Table 2), and
the average climate from available observations (Table 3).
Historically, Byrd and Ferrell are, respectively, the longest
and second-longest currently active AWS and were chosen
for their length of record. Siple Station also has an early start
date (1982) but was taken out of service in 1992 because of
logistical problems related to its location (remote from
McMurdo-based field support and in a very high accumulation rate region). Note that Siple Station AWS is far-removed
geographically from the Siple Coast and Siple Dome regions.
There are few AWS in the large region between Byrd and the
Antarctic Peninsula making the Siple Station record of
particular interest for understanding this area and its relationship to the rest of West Antarctica. The remaining three sites
were installed in 1986 and 1987 and chosen for this study to
fill out a rough transect between Siple Station and Ferrell.
Two geographic categories are apparent: ice sheet (Byrd and
Siple) and Ross Ice Shelf (the rest). Table 3 suggests possible
further climatic differences based on average pressure and
wind directions. All sites are within the south/southeast
Pacific sector of West Antarctica. Figure 2 summarizes the
availability of each AWS for the study period (1979 – 1993)
as the fraction of observations recorded each month. An
average for the suite of AWS gives a sense of how the ANNbased predictions enhance the overall climate record during
the study period. In general, availability of data at each AWS
is either quite high or quite low with few monthly values in
between. Siple’s outages in 1985– 1987 were not directly
related to failure of the meteorological instruments but to
other factors (primarily power-related; C. Stearns, University
of Wisconsin-Madison, personal communication, 2002).
Otherwise, most data loss is related to winter-season failures
and the subsequent delay until the austral summer field
season for repair. It is also worth noting that the available
pressure and temperature observations meet basic statistical
tests for normality (average skewness 0.2, average kurtosis
3.1) since predictive skill may be affected by the statistical
distribution of the training data.
2.2. ECMWF Data
[9] The ECMWF 15-year reanalysis product (ERA-15)
provides GCM-scale meteorological data for the period
1979 – 1993 [Gibson et al., 1999]. As a reanalysis product,
D04103
ERA-15 is based on a single version of ECMWF’s operational forecast model for the analysis period. Changes to the
model are thus removed as a source of (spurious) climate
change. Other factors, such as addition and removal of
observational data over time, remain as variables affecting
model skill, but these are external to the model. The original
ERA-15 production system used spectral T106 resolution
with 31 vertical hybrid levels. A lower resolution product
(used here) derived from those data provides 2.5 horizontal
resolution for the surface and 17 upper air pressure levels.
Six-hourly data are available at 0, 6, 12 and 18 UTC. A
subset of ERA-15 variables and vertical levels was used to
predict AWS temperature (Table 4) and pressure (Table 5).
In the horizontal, grid points adjacent to the AWS plus
points from the corners of a square area centered approximately on each station were used. Each grid point included
all selected variables. Reusch and Alley [2002] provide
further details on variable and grid point selection and
briefly review the validity of ERA-15 in Antarctica [see
also Bromwich et al., 1995; Cullather et al., 1997].
[10] ECMWF assimilates a number of Antarctic AWS into
its current operational forecasts, which opens the question of
circularity being present in this work (i.e., predicting observations for an AWS that’s been assimilated into the predictor
data). Identifying which sites were used in the ERA-15
project, and for which time periods, is extremely difficult
because of records not being readily accessible. However,
the impact of this issue is greatly reduced by using only
ERA-15 upper air data in the prediction process, particularly
since none of the AWS would have provided anything but
surface observations (i.e., no radiosonde data on the upper
atmosphere). Furthermore, for the AWS with the worst
records, it is unlikely that they would have passed the initial
data quality checks during the data assimilation process.
3. Methods
[11] This section describes our ANN-based methods, used
to create the complete 15-year records of AWS temperature
and pressure, and our method for estimating season length
Table 3. Climate Statistics Based on All Available AWS
Observationsa
Temperature,
C
Pressure,
mbar
Wind
Speed,
m/s
Station
nb
m
s
m
s
m
s
Siple
Byrd
Lettau
Elaine
Marilyn
Ferrell
20,608
41,882
30,786
9492
28,939
44,422
24.1
27.4
25.5
25.2
23.4
24.6
10.0
11.4
14.4
12.8
11.3
12.1
862.2
808.9
983.7
982.4
982.1
984.4
14.8
10.8
11.5
12.1
10.6
11.0
5.2
6.8
4.1
4.8
5.9
5.6
3.8
3.9
3.0
4.0
4.1
4.4
Wind
Directionc
A, deg L, %
191.6
4.7
167.2
177.3
257.3
215.6
35.6
82.6
45.7
38.5
60.6
52.5
a
Wind data are presented for climatic context only. Note the change in
mean wind azimuth A between Elaine and Marilyn.
b
Average number of available observations in the 3-hourly database.
Availability varies by instrument because of different operating characteristics (e.g., a temperature sensor is more robust than an anemometer).
c
The regular mean and standard deviation are not appropriate for wind
direction. Instead, a mean azimuth A and persistence L have been calculated
on the basis of the N/S and E/W directional components of the wind (V and
W, respectively)
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi [Middleton, 2000]. A = arctan (W/V) and L =
100* V 2 þ W 2 =n. L is a measure of variability around the mean azimuth.
3 of 28
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
D04103
D04103
Figure 2. Relative proportions of observations (gray) and predictions (white) in AWS temperature and
pressure records on a monthly basis. See Tables 1 – 3 for specific site data.
from the new temperature records. Results from the season
length analysis, and other climatological studies of the new
records, are presented in section 4 below.
Mexico [Hewitson and Crane, 1994b], prediction of summer rainfall over South Africa [Hastenrath et al., 1995] and
northeast Brazil [Hastenrath and Greischar, 1993], extreme
event analysis in the Texas/Mexico border region [Cavazos,
1999], and characterization of the monsoon climate of the
southwest United States [Cavazos et al., 2002]. We recommend these sources for a more detailed treatment of ANN
usage.
[13] ERA-15 surface data could be used directly to fill the
gaps in the AWS record. However, owing to the complexity
of near-surface conditions in Antarctica and to some fea-
3.1. Artificial Neural Networks
[12] Numerous sources for the theory and practical use of
artificial neural networks are available in the literature [e.g.,
Demuth and Beale, 2000; Gardner and Dorling, 1998;
Haykin, 1999; Hewitson and Crane, 1994a]. Examples in
meteorology and climate include developing an improved
understanding of controls on precipitation in southern
Table 4. Summary of ANN Predictor Variables for Temperaturea
Temperature,c C
Station
Time
Byrd
Elaine
Ferrell
Lettau
Marilyn
Siple
Y
Y
Y
Y
Y
Y
b
925
e
850
Temperature Advection,d C/km
700
925
850
Y
Y
Y
Y
Y
Y
700
600
Thickness,d m
850 – 700
f
850 – 500
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
a
700 – 500
Y
Y
Column headings refer to the predictor variable name and its pressure level, if appropriate. Thickness is based on the height difference between two
pressure levels with both being listed. A ‘‘Y’’ indicates that the variable was used as a predictor at that AWS site.
b
Day of the year divided by the total days in the year.
c
Extracted directly from ECMWF data sets.
d
Derived from ECMWF variables.
e
Pressure level in millibars.
f
Pressure levels used to compute layer thickness.
4 of 28
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
D04103
Table 5. Summary of Predictor Variables for Pressurea
Geopotential Height,b m
Station
Byrd
Elaine
Ferrell
Lettau
Marilyn
Siple
925d
Y
850
Y
Y
Y
Y
Thickness,c m
700
600
Y
Y
Y
Y
Y
850 – 700e
700 – 500
Y
Y
Y
Y
Y
Y
Y
Y
a
Details as in Table 4.
b
Extracted directly from ECMWF data sets.
c
Derived from ECMWF variables.
d
Pressure level in millibars.
e
Pressure levels used to compute layer thickness.
tures of ERA-15 (e.g., its assumption that the 500 m thick
Ross Ice Shelf is permanent, 2-m-thick sea ice, [Bromwich
et al., 2004]), surface data from ERA-15 (Figure 3) are not
as accurate as reconstructions for the free atmosphere.
Temperature predictions are significantly flawed in the
ERA-15 product (Figure 3a) while pressure predictions
are better but not without problems (Figure 3b). These
issues rule out use of ERA-15 surface data to fill the
temporal discontinuities in the high-quality AWS surface
observations. (A detailed comparison of ERA-15 skill
versus our ANN-based methodology is presented below in
section 5.) Surface conditions are nonetheless physically
related to conditions aloft, albeit through a poorly known
transfer function, which suggests that the latter may be
useful for empirical prediction of surface data. ANNs offer
arguably the most powerful techniques for empirically
determining how conditions aloft are translated to the
surface, thus making ERA-15 upper air data viable for
interpolation across data gaps in the AWS records. ANNs
are especially appropriate in a study such as ours in which
estimating surface meteorological conditions during AWS
data gaps is more important to us than is the transfer
function (physically how the free-atmosphere conditions
affect the surface).
[14] The ANN-based methodology used to develop the
records for the six AWS sites is essentially the same as
described by Reusch and Alley [2002]. Briefly, for each
AWS site and variable (temperature, pressure), an ANN was
trained to predict the AWS observations from selected
ECMWF variables at nearby grid locations with different
predictors used for temperature (Table 4) and pressure
(Table 5). Because of different elevations (Table 1) the
same pressure levels could not be used at all six sites, but
this did not affect the ANN training methodology. Instead,
ERA-15 data from pressure levels higher in the atmosphere
(e.g., 600 mbar versus 850 mbar) were used for the ice sheet
sites (Table 5). Because ANN skill is sensitive to initial
conditions and the order of training data, multiple ANNs
were trained from random initial settings with random
selection of training data before selecting the final version.
Standard ANN training uses a training subset of the input
data to adjust the ANN parameters (i.e., reduce prediction
error) and a testing subset to test predictive skill on data not
used in the adjustment phase [e.g., Haykin, 1999]. The most
skilled network has the lowest error on the training subset
but also performs well with the testing subset (i.e., the
network is still able to successfully generalize to new data).
D04103
All parameters having a significant impact on the predictive
skill of our ANNs were explored during the ANN training
process. For example, we used three types of feed-forward
ANN to find the most suitable architecture. Care was thus
taken to avoid overfitting, to test predictions against observational data not used in training, and to estimate the
reliability of the results, all following recommended modern
techniques in the field [Demuth and Beale, 2000; Haykin,
1999].
[15] Additionally, a number of refinements to the original
methodology were tested in the course of developing the
new records documented here. The ANN training was
originally based on one calendar year of AWS observations.
This approach was used primarily to demonstrate the
extreme case of having only one year’s worth of AWS data
available for ANN training, a situation likely to be true only
for the most recently installed AWSs. Assuming more data
are available, one might expect better skill when more data
are used for training. This did not turn out to be true as the
ANN predictive skill did not appear to change (positively or
negatively) appreciably with a larger training set. The other
significant refinement considered was the addition of insolation as another predictor variable for temperature. This
also led to little change in predictive skill most likely
because the annual cycle is already captured through use
of the Julian day predictor. Thus the issue of capturing the
complexity of surface air temperatures, due to phenomena
such as the strong inversion layer (most dominant in the
winter), remains unresolved. Other refinements may, of
course, improve these results (e.g., preprocessing the input
data to improve signal-to-noise ratios may be beneficial).
However, as our focus was on finding the simplest methodology first, and we have obtained useful results as is,
further modifications have been left for future work. As in
the previous work, we continue to use the Matlab Neural
Network Toolbox [Demuth and Beale, 2000; Haykin, 1999],
a widely available and well-documented implementation of
ANNs.
3.2. Season Length
[16] The length-of-season results reported in section 5 are
based on a somewhat subjective analysis of the complete
AWS temperature records that only attempted to identify
summer and winter seasonal conditions. Because these
seasons so dominate the temperature records, spring and
fall season lengths were not explicitly analyzed. They can,
however, be loosely estimated from the summer and winter
results. Analysis started with a 30-day running mean of
daily temperature values to reduce the short-term variability
in the record. The start and end of each season were
estimated using the mean and standard deviation of unsmoothed daily temperatures for that season (based on
climatological definitions of the seasons, i.e., DJF and
JJA for summer and winter, respectively). Transitions into
and out of a season were based on the temperature passing a
threshold value based on the seasonal mean plus or minus a
fraction of the seasonal standard deviation. For summer
(winter) transitions, the standard deviation was subtracted
(added) from (to) the seasonal mean to determine the
threshold value. To avoid problems with short-term intraseasonal changes confusing the threshold identifications,
transitions out of a season were considered tentative until
5 of 28
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
D04103
Figure 3. Scatterplots of ERA-15 predictions versus observations: (a) 2-m temperature and (b) 2-m
pressure. Data are shown as 6-hourly (lightest gray), daily (medium gray), and monthly (darkest gray)
values. Dashed lines offset from diagonal solid line represent ±RMS error from the correct (observed)
value.
the next season entry transition was found. With this extra
caveat, a short mid- or late-winter warming, for example,
that exceeded the winter exit threshold is not considered the
end of winter if temperatures return below the threshold
before summer temperatures are reached. Both the time step
size and offset were varied to attempt to improve the
robustness of the estimates. Step sizes ranged from daily
up to weekly (every seventh day). Offsets varied within the
range of the step size. In this way, season lengths were
computed over a range of temporal resolutions and subintervals. Results from each offset and step size combination were then averaged to produce the final season length
records. Even with these extra steps, the records retain a
certain amount of subjectivity and are shaped by our
assumptions on how to define seasonal transitions and the
length of record smoothing (30 days). Different choices
should produce only slightly different results even though
we have not exhausted the range of parameters that could be
used to identify season length.
4. Results
[17] Records for Ferrell were previously published by
Reusch and Alley [2002]. Scatterplots of ANN predictions
versus AWS observations are shown in Figure 4 at three
timescales (the original 6-hourly data and daily and monthly
averages). Corresponding statistics are presented in Table 6.
As expected, errors are reduced as averaging intervals
6 of 28
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
D04103
Figure 3. (continued)
increase. RMS values for monthly averages are significantly
improved versus the 6-hourly data. For example, the all-site
mean RMS values decrease by 4C and 1.9 mbar for
temperature and pressure, respectively. The scatterplots
(Figure 4) show (somewhat subjectively) that the predictions do not have any strong bias, i.e., points are uniformly
distributed around the 1:1 perfect prediction line and errors
of prediction are distributed fairly equally above and below
the target observations. Extreme values are offset in some
cases because of under/over prediction of maxima/minima.
There is more room for improvement in the temperature
predictions as compared to the pressure predictions. As
shown below in section 5, the results for temperature are
still significantly improved over ERA-15 surface data at the
nearest grid point to each AWS.
[18] Figure 5 presents the complete daily average records
for all six sites. As in the paper by Reusch and Alley [2002],
these records are composites made up of AWS observations
and ANN predictions for those time steps when observations were missing. An error envelope based on RMS errors
in prediction of observations appears around each section of
ANN prediction. For each year with observations, the RMS
error is from that year. For those years when no observations are available, e.g., before the AWS was installed, the
RMS error shown is an average of all RMS errors from
years with observations.
[19] To give a better sense of the ANN-based predictive
skill at 6-hourly resolution, two months with few missing
observations at all sites were selected for a more detailed
comparison of predictions to observations. (A monthly
timeframe lets us compare the most sites whereas fewer
sites are available at longer timeframes, e.g., seasonal,
because of more missing observations.) January 1988 (summer, Figure 6) is one of a small number of months with
nearly complete data at all six sites (Figure 2) but is not
otherwise intended to be representative. Plots for a winter
7 of 28
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
D04103
Figure 4. Scatterplots of ANN-based predictions versus AWS observations: (a) 2-m temperature and
(b) 2-m pressure. Colors and lines are as in Figure 3.
season month (June 1990, 5 of 6 sites available) are not
substantially different. Assessing the skill of the ANNbased temperature predictions (Figure 6a) is difficult
because of the higher frequencies present in the data.
Nonetheless, trends are generally matched in magnitude
and the errors are not strongly biased in either direction.
ANN-based pressure predictions (Figure 6b) track the
AWS observations quite well at all sites (in part because
pressure fields are more uniform). Synoptic changes (i.e.,
day-to-day or longer) of varying amplitudes are generally
reproduced with excellent fidelity. Where a mismatch is
present, it is usually not large and does not disrupt longterm trends. It should be noted that this analysis is not
strictly analogous to comparisons of climate model predictions and AWS observations [e.g., Cassano et al.,
2001]. Climate model predictions are physically based
and dependent on values from previous time steps.
ANN-based predictions derive from empirical relationships
discovered during training and are not serially correlated
since no time lags are used in the predictor data.
5. Discussion
[20] This section begins by placing our prediction results
in context with ERA-15 surface predictions. Aspects of
temporal and spatial variability in the composite temperature and pressure records are then discussed followed by
results from our length-of-season analysis. Additional variability topics (e.g., EOF analysis) are covered by Reusch
[2003].
5.1. Comparison With ERA-15 Surface Predictions
[21] As mentioned under Methods, ECMWF surface data
could also be used to fill the gaps in the AWS records and,
8 of 28
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
D04103
D04103
Figure 4. (continued)
Table 6. ANN Prediction Statisticsa
Temperature
6-hourly
Pressure
Daily
Monthly
6-hourly
Daily
Monthly
Station
n
RMS
r
n
RMS
r
n
RMS
r
n
RMS
r
n
RMS
r
n
RMS
r
Byrd
Elaine
Ferrell
Lettau
Marilyn
Siple
Mean
14,739
5107
16,281
10,305
7413
10,496
7.1
6.4
6.6
8.1
5.7
5.3
6.5
0.79
0.88
0.83
0.83
0.86
0.85
3731
1286
4156
2611
1909
2650
6.1
5.4
5.7
6.8
4.5
4.3
5.5
0.83
0.91
0.88
0.87
0.90
0.89
127
45
144
90
70
88
3.0
2.2
2.8
3.4
2.0
1.8
2.5
0.95
0.98
0.97
0.96
0.98
0.97
16,051
5170
16,305
10,303
8025
10,496
4.1
3.9
3.3
3.9
3.8
3.7
3.8
0.93
0.95
0.96
0.94
0.95
0.94
4071
1302
4157
2611
2065
2650
3.8
3.4
2.9
3.4
3.5
3.4
3.4
0.94
0.96
0.96
0.95
0.95
0.95
139
45
144
90
75
88
1.9
2.0
1.5
1.4
3.1
1.5
1.9
0.98
0.97
0.97
0.98
0.91
0.97
a
RMS errors and correlation coefficients are shown for comparisons with 6-hourly, daily, and monthly observations (averaged on daily and monthly
timescales). Here, n is the number of AWS observations available. The mean RMS error is across the six AWS sites.
9 of 28
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
Figure 5. Daily average records from AWS sites of this study: (a) 2-m temperature and (b) 2-m
pressure. Observations are shown as a blue line, and ANN predictions are shown as a red line.
Confidence intervals derived from RMS errors appear as a light gray envelope around predictions. See
color version of this figure at back of this issue.
10 of 28
D04103
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
Figure 5. (continued)
11 of 28
D04103
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
Figure 6. Detailed comparisons of ANN-based predictions (gray) and AWS observations (black) for
January 1988 for (a) 2-m temperature and (b) 2-m pressure.
12 of 28
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
D04103
D04103
Table 7. ERA-15 Prediction Statisticsa
Pressureb
Temperature
6-hourly
Daily
Monthly
6-hourly
Daily
Monthly
Station
n
RMS
r
n
RMS
r
n
RMS
r
n
RMS
r
n
RMS
r
n
RMS
r
Byrd
Elaine
Ferrell
Lettau
Marilyn
Siple
Mean
14,739
5107
16,281
10,305
7413
10,496
8.2
10.9
10.6
13.6
5.2
6.9
9.2
0.84
0.70
0.88
0.89
0.88
0.79
3731
1286
4156
2611
1909
2650
7.7
10.4
10.3
13.3
4.4
6.2
8.7
0.87
0.72
0.91
0.91
0.91
0.82
127
45
144
90
70
88
5.9
7.3
9.4
11.9
2.0
3.2
6.6
0.95
0.87
0.96
0.98
0.98
0.95
16,051
5170
16,305
10,303
8025
10,496
16.3
11.8
3.1
4.3
3.4
7.3
7.7
0.85
0.51
0.97
0.95
0.95
0.91
4071
1302
4157
2611
2065
2650
16.2
11.4
2.8
4.0
3.2
7.2
7.5
0.85
0.51
0.98
0.96
0.96
0.91
139
45
144
90
75
88
15.4
6.6
2.3
3.2
3.1
6.3
6.2
0.82
0.57
0.96
0.94
0.88
0.92
a
RMS errors and correlation coefficients are shown for comparisons with 6-hourly, daily, and monthly observations (averaged on daily and monthly
timescales). Here, n is the number of AWS observations available. The mean RMS error is across the six AWS sites.
b
RMS values for pressure may be inflated for higher-elevation sites (Byrd, Siple) because of errors in reducing AWS observations to sea level. We used
the hydrostatic equation P = rgz/100 to determine the adjustment (in millibars), with r = 1.29 kg/m3 and g = 9.8 m/s2.
in theory, should be just as reasonable as any data from an
empirical methodology (e.g., the satellite-based methodology of Shuman and Stearns [2001]). They have the advantages of being available at all 6-hourly time steps and
serving as an alternative benchmark since these data were
not used in the ANN training process. Scatterplots of ERA15 surface predictions versus AWS observations are shown
in Figure 3. Corresponding statistics are presented in
Table 7. ERA-15 surface temperatures have obvious problems at the four ice shelf sites and minor problems at the
two ice sheet sites. Lettau, Elaine and Ferrell temperatures
are strongly biased and feature a floor of minimum values
around 30 to 35C. The obvious candidate explanation
for these errors in the minima is ERA-15’s treatment of the
ice shelf as permanent, 2-m thick sea ice, which will cause
problems with surface heat fluxes between the ocean and
atmosphere [Bromwich et al., 2000, 2004] since the ice
shelf is in fact closer to 500 m thick. This alone will lead to
anomalously high surface temperatures, particularly during
winter when the ocean-atmosphere temperature gradient is
strongest. Temperatures at Marilyn are also biased but to a
lesser extent than the other ice shelf sites. Marilyn’s
proximity to grounded ice and the relatively low resolution
of the ERA-15 land-sea mask (possibly placing Marilyn on
land instead of the ice shelf) may explain the reduced errors
at this site. Problems with Byrd temperatures suggest that
ERA-15 has a substantial cold bias at temperatures below
approximately 30C. Overall, the ERA-15 surface temperatures are much less accurate compared to our ANN
predictions. Even for monthly averages, the ERA-15 average RMS error for all AWS sites is 6.6C (versus 2.5C for
the ANNs).
[22] An alternative for temperature-only comparisons is
the satellite-based temperature reconstructions of Shuman
and Stearns [2001]. Three of our AWS overlap with this
study. Standard deviations of the difference between daily
mean observed and their predicted temperatures were
6.25C, 5.67C and 5.22C for Byrd, Lettau and Siple,
respectively [Shuman and Stearns, 2001]. Similar ERA-15
standard deviations are 6.3C, 7.0C and 6.3C, while our
ANN prediction statistics are 6.1C, 6.7C and 4.3C.
Results for the three approaches vary from site to site with
the largest difference in skill seen with Siple. A more
detailed comparison of our results with Shuman and Stearns
[2001] is presented by Reusch and Alley [2002] for Ferrell
AWS.
[23] Pressure predictions by ERA-15 are generally much
better than for temperatures (Figure 3b and Table 7). The ice
shelf sites, excluding Elaine, are comparable to the ANN
predictions. Determining the skill for the ice sheet sites is
more complex because of the need to reduce the AWS
observations to sea level so that they can be directly
compared to the ERA-15 values. Thus some of the reported
RMS error may be due to how we adjusted the AWS
observations (see Table 7 for details). ERA-15 pressure
errors at Elaine are substantial (too high and too low) albeit
relatively unbiased. The monthly RMS value of 6.6 mbar is
more than twice that of Lettau (3.2 mbar), its nearest
neighbor. Combined with the temperature errors, the
ERA-15 data for Elaine are remarkably poor. In contrast,
monthly errors for Marilyn (2C, 3.1 mbar) are quite similar
to our ANN results. It should not be forgotten, however, that
we can only judge the quality of ERA-15 surface data when
there are AWS data available for comparison. Thus it is not
safe to assume, on the basis of these comparisons, that all
ice sheet grid points will have acceptable ERA-15 surface
temperatures. On the basis of our analysis and the known
model limitations, it is likely, however, that temperature
errors in the ERA-15 data will not be acceptable at most ice
shelf grid points.
[24] Figure 7 compares the skill of our ANN-based
predictions to the ERA-15 surface predictions with monthly
average RMS values. With few exceptions, the ANN-based
predictions have higher skill, substantially higher in many
cases. As seen in the paper by Reusch and Alley [2002],
there are generally no strong patterns in the ANN-based
prediction errors. Minor exceptions can be seen at Marilyn
and Siple where the pressure errors (Figure 7b, black bars)
tend to peak in the austral winter months. ERA-15 temperature errors (Figure 7a, white bars) have substantial monthly
to seasonal-scale variability at all sites with no obvious
spatial pattern. A strong annual cycle appears in the ERA-
Figure 7. Average monthly RMS error of ANN-based (black) and ERA-15 (white) surface predictions versus AWS
observations: (a) 2-m temperature and (b) 2-m pressure. Values for pressure may be inflated for higher-elevation sites
(Byrd, Siple) because of errors in reducing AWS observations to sea level.
13 of 28
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
Figure 7.
14 of 28
D04103
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
15 pressure errors for Siple and Byrd (Figure 7b, white
bars). RMS errors are much lower during the winter months.
As noted earlier, this cycle may be partly due to the
elevation of these two sites (Table 1) and the process of
reducing observed pressures to sea level for comparison. It
is unlikely, however, that this would explain the observed
cyclicity.
5.2. Temperature Variability
5.2.1. Overview
[25] Qualitative (Figure 8) austral winter-centered overviews of the variability in the AWS temperature records
readily show seasonal cycles, interannual variability and
significant intersite differences. The subjective sense of an
underlying similarity between sites is supported by twodimensional correlations of the monthly data in Figure 8 (R >
0.9 for all comparisons). Features common to all sites are
the steep transitions into and out of winter, with fall and
spring being relatively short seasons. Figure 8 also highlights the variability in depth and length (start/end) of the
winter season. Winter conditions (coldest temperatures)
tend to exist nearly half the year in many cases. These are
all characteristics of the ‘‘kernlose,’’ or ‘‘coreless,’’ winter
[e.g., van Loon, 1967; Wexler, 1958]. Polar continental
winter temperatures, e.g., in Siberia, typically have a sharp
minimum value [e.g., Wexler, 1958, Figure 2]. In contrast,
the kernlose winter has a broad, flat winter minimum. Many
Antarctic stations also often have short-term early winter
reversals of the expected cooling trend due to warm air
advection associated with cyclonic activity [van Loon,
1967; Wexler, 1958]. An austral-summer-centered version
of Figure 8 (not shown) suggests a modest summer warming trend at Siple, Lettau and Marilyn and, possibly, a
period (1989 – 1992) of relatively cooler summers at Byrd.
Ferrell has the longest summer season and the least interannual variability. Ferrell also shows above-average December temperatures for 1987– 1991. Reference to Figure 4
suggests that this warming may be due simply to a cold bias
in the ANN-based predictions, i.e., the earlier part of the
record is dominated by ANN-based predictions that are too
cold. While this explanation cannot be entirely ruled out, it
is not well supported by comparisons of predictions and
available observations (Figure 4), which do not show a
significant bias. Predictive skill is known to vary somewhat
by season at Ferrell [Reusch and Alley, 2002] but it is the
best during the summer. Thus the Ferrell warming in the
latter half of the study period likely has a real physical basis.
5.2.2. Temperature Trends
[26] A thorough linear regression analysis of the complete
temperature time series, including sliding windows of
varying lengths, reveals essentially no statistically significant trends at all six sites based on the regression slopes.
(Statistical significance based on a = 0.05 confidence
intervals.) Lettau does have a warming trend (0.06C/yr)
in 6-hourly data but it disappears over longer averaging
periods. This value is also well below the known instrument
accuracy level (0.25 – 0.5C). When subannual periods are
examined on a year-to-year basis (e.g., winter-to-winter),
statistically significant warming trends appear at three sites
during the austral summer season. Siple, Lettau and Marilyn
have summer season warming trends of 0.18– 0.25C/yr.
At monthly resolution, warming is present in January at all
D04103
3 sites (0.25– 0.27C/yr) but only at Siple and Marilyn in
December (0.27 and 0.3C/yr) and February (0.14 and
0.26C/yr). Marilyn also shows warming in October and
November (0.31 and 0.48C/yr, respectively). The explanation for the directions and locations of these trends
remains speculative. The distance between Siple and the
two ice shelf sites (1777 and 2221 km for Lettau and
Marilyn, respectively) and the lack of a trend at Byrd
suggests there may be different processes involved. And
while Lettau and Marilyn are both on the Ross Ice Shelf,
they are 450 km apart and in potentially quite different
geographical conditions. For example, there is a substantial
shift (80) in the annual mean wind direction (Table 3)
between Lettau/Elaine and Marilyn that may account for
some of the differences seen between these sites. The shift
in the mean wind at Marilyn is related to the prevalence of
katabatic winds from the Byrd Glacier [Breckenridge et al.,
1993]. While katabatic winds are often considered a predominantly winter season phenomenon, seasonal mean
wind direction (259.2 and 254.7 for winter and summer,
respectively) suggests a year-round influence at Marilyn. As
measured by persistence (defined in Table 3), however,
directional variability is higher in the summer (40%, where
100% indicates a constant value) versus winter (73%).
Regardless of a fully satisfying physical explanation, it
appears that summer temperatures have been increasing
over time at three distinct West Antarctic AWS sites.
5.2.3. Temperature Correlations
[27] To assess similarities and differences between the
temperature records, correlations at varying temporal resolution (6-hourly to annual) were done between all the sites
(a total of 15 at each timescale). Seasonal and monthly
correlations are generally high (R > 0.95 and R > 0.91,
respectively). Yearly R values are more variable and lower
by 0.14 to 0.52 (mean 0.31). Average values, by site, for the
yearly correlations range from 0.57 (Siple) to 0.71 (Elaine).
The annual averages also suggest possible behavioral differences between the ice sheet and ice shelf sites. For the ice
shelf sites (Figure 1), the average intersite correlation is 0.76
with respect to the other ice shelf sites but only 0.58 with
respect to the two ice sheet sites. Siple and Byrd correlate
with each other at R = 0.67, slightly better than with the ice
shelf sites. On an interannual basis, seasonal correlations
show strongest intersite coherence in fall and winter (mean
R 0.7) and a substantially lower correlation in summer
(mean R = 0.44). Furthermore, during summer all sites
appear to become ‘‘disconnected’’ from Byrd (R < 0.27,
mean R = 0.09).
[28] To test dependence of the yearly correlations on the
calendar used to form the averages, we repeated the correlations using different monthly calendar boundaries. This
resulted in 11 new ‘‘annual’’ (i.e., 12-month but not
calendar year) correlations for each of the 15 intersite
comparisons. Except for Siple-Byrd, at least eight of the
alternate calendars produced higher intersite correlations. In
eight of the 15 comparisons, R values improved for all
alternative calendars. With a December – November calendar, the R values increased to 0.89 or higher. The weakest
correlation on the standard calendar (Siple-Lettau, R = 0.45)
exceeded 0.93 on the October – September, November –
October and December – November calendars. Similar
improvements are also seen with Siple correlations to the
15 of 28
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
Figure 8. Monthly mean temperatures (C) in a month-by-year format, centered on austral winter
(January to December, x axis). Years on y axis refer to January. Values are interpolated in x and y in order
to improve the rendering; thus the figure should not be taken absolutely literally at the smaller scales. The
center of the color bar has been set to white to highlight values above/below the mean. See color version
of this figure at back of this issue.
16 of 28
D04103
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
D04103
Figure 9. Austral winter-centered annual anomalies of temperature (C). Anomalies are with respect to
the 15-year average and values above (below) 1 positive (negative) standard deviation are shown in dark
gray.
other ice shelf sites (R values improve from 0.51– 0.68 to
over 0.93). Results between Byrd and the ice shelf sites also
show modest improvements (mean 0.15). The largest difference is between the normal and December –November
calendars, suggesting sensitivity to which December is
included, the most recent or the one 11 months in the
future. October and November also appear to play a role
on the basis of the improvements seen when starting the
calendar in these months. Changing calendar boundaries
can clearly improve intersite correlations but a mechanism
to explain why some calendars are better than others is
elusive. Nonetheless, studies of yearly averages need to be
aware of effects related to calendar selection.
5.2.4. Temperature Anomalies
[29] Apparent anomalies seen at many sites (Figure 8) are
warm winters in 1980 and 1988, cold falls in 1982 and
1987, and the lack of fully established winter conditions in
1987 and 1988 (particularly at Byrd). Austral-winter-centered annual anomalies (Figure 9) confirm warmth at all
sites in 1980, cold two (three) years later at the ice shelf (ice
sheet) sites, and a second warm anomaly at all sites but
Siple in 1988. Values range from 2.3C up to 2.6C
(Table 8). These values, and their timing, are subject to
change based on the bounds selected for the annual average.
For example, austral-summer-centered anomalies (not
shown) may produce a different view of interannual variability. Summer-centered anomalies shift the 1980 warm anomaly to 1981 suggesting that temperatures were abnormal
during the austral winter-spring of 1980. Annual anomalies
are still useful but subannual data provide a better picture of
the spatial and temporal pattern of variability at our sites.
Seasonal temperature anomalies (Figure 10) range from
6.1C up to 7.8C. The largest negative anomalies tend to
occur in fall while the largest positive anomalies occur in
winter (thus possibly biasing annual anomaly calculations).
Monthly temperature anomalies (Figure 11) range from
12.5C up to 10.9C. Extreme anomaly values, by site,
occur in September (minimum) and July/August (maximum).
5.2.4.1. Warm Events
[30] The 1980 warm anomaly is substantial at nearly all
sites and exceeds 2C at Byrd, Lettau and Elaine (Figure 9).
Figure 10 shows that the anomaly is predominantly due to
abnormal warmth in the austral winter of that year with
additional contributions from a warm fall. The evolution of
this anomaly is fairly similar at all sites. The 1980 warm
anomaly appears as high monthly values for July/August on
17 of 28
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
D04103
D04103
Table 8. Annual Temperature Statisticsa
Siple
Byrd
Lettau
Elaine
Marilyn
Ferrell
m
A
m
A
m
A
m
A
m
A
m
A
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
24.4
23.1
25.6
25.2
26.1
23.9
24.7
26.3
25.6
24.1
24.4
24.4
24.1
25.3
24.4
0.4
1.7
0.8
0.4
1.3
0.9
0.1
1.5
0.8
0.7
0.4
0.4
0.7
0.5
0.4
27.0
24.5
28.1
28.2
29.3
26.4
27.4
28.4
26.7
24.7
26.3
29.4
26.9
26.6
26.5
0.1
2.6
1.0
1.1
2.2
0.7
0.3
1.3
0.4
2.4
0.8
2.3
0.2
0.5
0.6
28.0
24.3
27.2
28.6
27.6
27.3
28.9
27.2
26.8
24.7
26.5
27.4
25.3
25.9
27.4
1.1
2.6
0.3
1.7
0.7
0.4
2.0
0.3
0.1
2.2
0.4
0.5
1.6
1.0
0.5
24.1
21.8
24.3
25.8
24.2
23.7
25.2
24.9
25.0
22.6
23.5
24.0
24.4
25.1
23.8
0.1
2.4
0.1
1.6
0.0
0.5
1.0
0.7
0.8
1.6
0.7
0.2
0.2
0.9
0.4
23.9
21.6
23.7
24.9
23.9
23.2
24.5
24.2
23.3
21.7
23.6
22.3
23.1
23.7
24.7
0.4
1.9
0.2
1.4
0.4
0.3
1.0
0.7
0.2
1.8
0.1
1.2
0.4
0.2
1.2
24.0
22.4
25.0
26.3
24.2
24.0
26.4
24.9
24.0
22.1
23.4
25.3
22.5
24.9
24.3
0.2
1.8
0.8
2.1
0.0
0.2
2.2
0.7
0.2
2.1
0.8
1.1
1.7
0.7
0.1
Mean
Std Dev
Min
Max
24.8
0.9
26.3
23.1
1.5
1.7
27.1
1.5
29.4
24.5
2.3
2.6
26.9
1.3
28.9
24.3
2.0
2.6
24.2
1.0
25.8
21.8
1.6
2.4
23.5
1.0
24.9
21.6
1.4
1.9
24.3
1.3
26.4
22.1
2.2
2.1
Year
a
Temperatures are given in degrees Celsius; m is the annual mean value, and A is the annual anomaly of the mean value with respect to the full record
mean. The latter is listed in the lower section of the table along with other statistics for the annual values.
Figure 10. Seasonal temperature anomalies. Anomalies are with respect to the 15-year seasonal
averages. Shading is as in Figure 9. Austral seasons are shown in the order summer, fall, winter, spring
and are based on the standard climate periods (DJF, MAM, etc.).
18 of 28
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
D04103
Figure 11. Monthly temperature anomalies. Anomalies are with respect to the 15-year monthly
averages. Shading is as in Figure 9. Months are in standard calendar order, January to December.
the ice sheet and July – September on the ice shelf. May (and
sometimes April) 1980 is also anomalously warm though
not always above one standard deviation. This monthly
pattern readily explains the seasonal signature of this warm
event.
[31] The 1988 warm anomaly also exceeds 2C at Byrd,
Lettau and Ferrell. This anomaly is also related to abnormal winter warmth (relative to the 15 year record) but in
this case the additional contributions come from a warm
spring. Only Siple does not show a strong (more than one
standard deviation) warm anomaly in this timeframe. The
spatial pattern is less coherent for the 1988 warm anomaly.
Siple has only a single (very) warm month (August),
which averages out on the longer timescales. Above
average warmth occurs at Byrd from June to October with
the strongest anomaly in August. Lettau also has a strong
anomaly in August plus another in October. The remaining
ice shelf sites all have anomalies from August to October.
Thus the 1988 event appears to have influenced Siple
differently from the other sites (only one warm month)
while having the strongest effect on the ice shelf sites
(with Lettau having additional influences). The four ice
shelf sites (Figure 1) also have a strong single-season (fall)
warm anomaly in 1983. It does not appear in the annual
anomalies because of cold anomaly offsets in other sea-
sons. This ice shelf-only anomaly is seen at all sites in
April/May.
5.2.4.2. Cold Events
[32] At the annual level (Figure 9), the main cold anomalies appear at the ice shelf sites in 1982 and at the ice sheet
sites in 1983. Seasonal data suggest that these are two
distinct events and the 1982 anomaly is actually present at
all sites. The 1982 annual averages are cold because of a
cold fall at all sites (Figure 10). This anomaly tends to be
largest in April and is present from two to five (Ferrell)
months (Figure 11). It is not seen in the ice sheet annual
averages because of masking by offsetting warm/neutral
seasons.
[33] The 1983 annual cold anomaly at the ice sheet sites
comes from a cold winter, also seen at Marilyn (but masked
in the annual average by a warm fall). These three sites also
share a cold summer in 1981. This implies different behaviors, in at least some situations, between the ice sheet and
ice shelf sites. Furthermore, Marilyn also appears to have
characteristics of both geographic domains (also suggested
by the ERA-15 prediction statistics, Figure 3). A connection
between this cold anomaly and the 1982 – 1983 El Niño is
suggested by the link previously seen between the Southern
Oscillation Index (SOI) and temperatures at South Pole and
stations across West Antarctica [Savage et al., 1988]. The
19 of 28
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
SOI is an index of the atmospheric component of the
coupled ocean-atmosphere phenomenon known as ENSO
(El Niño/Southern Oscillation). Savage et al. [1988] found
that the extreme negative annual anomaly (base period
1957 – 1985) in the SOI in 1982 was followed a year later
by an extreme negative annual temperature anomaly (base
period 1957 – 1986) at South Pole. They also found that
monthly temperatures across West Antarctica for 1983 were
significantly correlated with South Pole, unlike all other
years in their study period (1980 – 1986). Two of our
stations overlap (Byrd and Siple) suggesting strongly that
the cold anomaly seen at our four other AWS is also tied to
the 1982 El Niño. Further investigation of links between our
temperature and pressure records and ENSO is planned.
[34] Other widespread seasonal cold anomalies appear in
1981 (spring, weak at Byrd), 1986 (winter-spring, winter
only at Siple), and 1987 (fall, weak at Ferrell). In most
cases, the spring 1981 anomaly occurs only in November
while stretching into December at Siple and Lettau. The
winter-spring 1986 cold anomaly peaks in September at all
sites, quite dramatically at most sites (greater than 10C
anomaly). August is also generally a substantial contributor
to this anomaly. The 1987 fall cold anomaly peaks in
March/April and is limited to these two months (unlike
the fall 1982 event).
5.2.4.3. Byrd and the 1987 – 1989 Winters
[35] The period 1987– 1989 is unique within the 15-year
record at Byrd with normal winter conditions apparently
never being fully established (Figure 8). Average monthly
winter temperature anomalies (Figure 11 and Table 9) are
remarkably high, particularly for 1987– 1988 and August
1988. This behavior first appears in 1987 and is best
established in 1988 when it extends into September. By
1989, only July remains significantly above normal. While
not all the values in Table 9 exceed one standard deviation,
taken together seasonally they lead to extended periods of
above average winter temperatures during 1987 and 1988.
5.3. Pressure Variability
5.3.1. Overview
[36] Austral winter-centered overviews (Figure 12) of the
variability in the AWS pressure records do not have the
well-structured seasonal cycle seen in the temperature data
(Figure 8). Instead, two ‘‘halves’’ of the year can be
distinguished on the basis of generally higher versus lower
pressures (December – June and July – November, respectively) but there is large interannual variability in the
magnitude and timing of the anomalies. The first half of
the year (December – June) can also be split into a consistently higher period (December – March) and an average
period (April – June) with broad positive anomalies in the
second half of the record. Departures from normally lower
pressures in the second half of the year occur in the early
record (1979 – 1981) and in the late 1980s. Overall, the sites
are broadly very much alike with the ice shelf sites having
the strongest similarities.
5.3.2. Pressure Trends
[37] Regressions through the complete pressure records at
various temporal resolutions identified just one site with a
significant trend for the full period. AWS Marilyn had a
small positive trend of 0.2 mbar per year with significance
at the 95% confidence level that disappears at monthly and
D04103
Table 9. Monthly Mean Temperature Anomalies at Byrd, 1986 –
1990a
Period
May
June
July
August
September
1986
1987
1988
1989
1990
1987 – 1989
1987 – 1988
Std Dev
3.2
1.1
0.7
4.2
1.8
2.0
0.9
2.9
4.2
3.9
4.0
0.1
7.0
2.6
4.0
3.9
1.7
6.0
3.8
3.3
3.0
4.4
4.9
4.6
4.7
3.7
9.6
0.6
5.8
4.6
6.6
5.3
10.5
0.0
5.2
0.9
2.8
2.0
2.6
3.6
a
Values are given in degrees Celsius and are relative to full record
monthly average. Standard deviation is calculated for full 1979 – 1993
period.
longer intervals. As this is approximately the same as the
instrument calibration (±0.2 mbar) and resolution
(0.05 mbar), it is not clear that the trend has physical
significance. Sliding window regressions of the annual
average data identify significant trends at Byrd, Marilyn
and Ferrell above what might be expected by chance, but
the larger trends are only found in the shorter window
lengths (i.e., 3 and 4 years). Window lengths of 9 to 12
years have significant but small trends for the middle period
of the record (including the first and last one to two years
removes the significance). Thus, as strongly suggested by
the time series (Figure 5b), the statistics do not appear to
support any substantial trends in near-surface pressure over
the study period apart from short-term fluctuations. Significant trends are also absent from the interannual time series
of seasonal and monthly averages.
5.3.3. Pressure Correlations
[38] Full-record correlations with annual, seasonal and
monthly averages give generally high R values with correlations improving as the averaging period decreases. R
values are nearly all 0.95 or better among the ice shelf
sites. Correlations between the ice sheet and ice shelf sites
are mostly from 0.7 to 0.9. Siple disconnects from Marilyn
at the annual scale with R = 0.46. Interseasonal correlations
are nearly all significant and high (mean R = 0.85) except
for Siple and the ice shelf sites during winter, which are low
(R < 0.4) and nonsignificant. Intermonthly correlations are
also nearly all high and significant with Siple and Byrd
being the exceptions with low correlations for August. The
latter likely explains Siple’s low correlation to the ice shelf
in the winter season.
[39] Further analysis of the annual pressure data with
sliding window correlations produces three site groupings.
The first group is the 6 correlations involving the 4 ice shelf
sites. These sites have high correlations (average R = 0.9)
and statistical significance for nearly all window lengths and
intervals. This is likely explained by the close proximity of
these sites (average intersite distance is 400 km, Table 2)
and the relatively homogeneous nature of the ice shelf
environment on annual timescales. (The latter is due in part
to the moderating influence of the nearby marine environment.) The two ice sheet sites, Siple and Byrd, separately
form the second and third groups. Byrd correlates well with
both Siple and the ice shelf sites whereas Siple’s correlations with the ice shelf are substantially reduced. The
increased distance from the ice shelf to Siple relative to
Byrd, over 900 km, is the likeliest explanation. The Siple-
20 of 28
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
Figure 12. As in Figure 8 but for monthly mean pressure. See color version of this figure at back of this
issue.
21 of 28
D04103
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
D04103
Figure 13. As in Figure 9 but for annual mean pressures.
to-ice-shelf relationship does have a modest temporal aspect
in that for 3 of the sites (excluding Marilyn) there tend to
be stronger correlations for windows that include the first
half of the record. That is, R values decline as the later
part of the record is included in the correlation window,
particularly for the shorter windows. This implies that
there may have been less spatial variability in the earlier
part of the study period (up to 1985 or 1986, approximately). Given that Marilyn does not correlate at all (R 0)
with Siple in these analyses, overinterpreting this apparent
trend may be premature.
5.3.4. Pressure Anomalies
[40] As with temperature, significant anomalies (greater
than one standard deviation) exist in the pressure records.
At the annual scale (Figure 13) there is more intersite
variability than was seen in the temperature records
(Figure 9). Only one positive anomaly (1980) occurs at
more than two sites (Siple, Byrd and Lettau; weak at
Ferrell). The four ice shelf sites tend to share the same
pattern of anomalies but at varying magnitudes. Seasonal
(Figure 14) and monthly (Figure 15) anomalies show a
higher level of spatial similarity. All sites now have positive
anomalies in the spring of 1980 and 1988, the fall of 1990,
and the summer and winter of 1992. Similarly, negative
anomalies are present at all sites in the fall of 1982 and the
spring of 1986. The four ice shelf sites have strong negative
anomalies in winter 1993. The seasonal anomalies tend to
have the same expression at each site in the monthly values,
but the latter also show a number of strong anomalies too
short to appear in the longer averages. For example, May
1989 has a negative departure of 15 mbar or more at all
sites. Many of the pressure anomalies are associated with
the temperature anomalies described earlier but the relationship is not perfect. For example, the winter warm anomaly
of 1980 (Figure 10) is not seen in the ice shelf pressure
anomalies (Figure 14). Conversely, the four seasons of
positive pressure anomalies at Byrd in the same year are
only seen in the fall and winter temperature anomalies. Not
surprisingly, the two meteorological variables are not in
lockstep.
[41] Figure 16 provides a context for four of the seasonal
pressure anomalies. The positive anomalies of spring 1980
and 1988 (Figures 16a and 16b) are associated with a
relatively shallow Amundsen Sea Low (ASL) centered
around 110 – 130W. The area of higher pressure (more
than 996 mbar) on the plateau is also quite large (though
it is important to remember that these values have been
reduced to sea level from elevations up to 4 km). The two
seasons shown with negative pressure anomalies, spring
1986 and fall 1982 (Figures 16c and 16d), each have a
deeper ASL but in quite different positions. In the spring of
1986, the ASL has shifted westward to 150W and thus is
22 of 28
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
D04103
Figure 14. As in Figure 10 but for seasonal mean pressures.
partly over the Ross Sea. In contrast, the center of the
ASL has shifted eastward to 90W in the fall of 1982
although significant low pressure still extends past 170W.
The low-pressure center in the eastern hemisphere is also
located quite differently between these two seasons.
Higher pressure over the plateau is greatly reduced in area
in both these seasons relative to the two positive anomaly
seasons.
5.4. Length of Season
[42] Results from our temperature-based analysis of
winter and summer season lengths (Figure 17) show the
winter season mean to be roughly twice as long as the
summer season (note the different axes for the two
seasons), matching what is subjectively seen in Figure 5.
The time series plots (Figure 17) show how season lengths
have changed over the study period. The importance of
season length is primarily related to issues of energy
balance. For example, more ice might melt, or ice melting
might be started, by a longer summer. Season length is of
great importance in coastal areas where sea ice characteristics are most likely to be affected by variations in
seasonal warming and cooling. Sea ice variability, in turn,
can have biological and oceanographic effects. Coastal ice
core records of climate may also be affected [Wolff et al.,
1998]. Although there are a number of apparent trends in
season length, particularly in the summer data, the only
statistically significant trend found with robust regression
is at Lettau. The summer season increases by 2.5 days
per year over the study period at this site. A comparison of
the two seasons shows interannual variability to be larger
during winter. This is particularly noticeable at Elaine,
Marilyn and Ferrell where the site-average standard deviations are 1.3 and 4.1 weeks for summer and winter,
respectively. Also of note in the time series is the
interannual coherency across most or all of the sites,
particularly during the winter. For example, the first five
years of the winter records are highly correlated. Only in
the last few years of the record does this pattern break
down completely. The summer season records also show
coherent periods but they are not as well defined.
6. Conclusions
[43] Application of our artificial-neural-network-based
methodology for AWS record enhancement to a site on
the Ross Ice Shelf (Ferrell) given by Reusch and Alley
[2002] demonstrated the utility and skill of this technique
for the period 1979 – 1993. The work presented here
expands the ANN-based approach both to other sites on
the Ross Ice Shelf and to sites at elevation (1000 m and
higher) on the West Antarctic Ice Sheet. In total we have
23 of 28
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
D04103
Figure 15. As in Figure 11 but for monthly mean pressures.
produced complete 15-year records of temperature and
pressure for six AWS sites across West Antarctica. That
the skill of this technique is comparable at all sites, both
quantitatively (e.g., RMSE, R2) and qualitatively (e.g.,
scatterplots), suggests that at least some of the residual
predictive error is due to unsolved problems with the
methodology and is not entirely due to site-specific details.
For AWS temperature, however, there appear to be aspects
of the upper air-surface relationship (e.g., the surface
inversion) that require new approaches to this problem to
be considered. It is worth noting that the ANN-based skill
for AWS temperatures is still better than that of a state-ofthe-art mesoscale forecast model and without seasonal bias
(i.e., the Polar MM5 [Bromwich et al., 2004]). Planned
future work will investigate a number of new possibilities
(e.g., preprocessing of the input data to improve the signalto-noise ratio) and expand the set of AWS sites. Unfortunately, there are no other sites in the large region between
Siple Station and Byrd where additional data would be
particularly useful to ice-coring activities in this area. New
reanalysis products (ERA-40) offer the opportunity to
extend the AWS records further back in time at the expense
of recalibrating the ANNs.
[44] Results from climatological analysis of these records
serve to confirm many previous inferences based on different and often less-consistent data sets, strengthening many
prior conclusions about West Antarctic weather and climate. Major new insights have not emerged from the new
analyses, either because all of the major phenomena have
been identified already, or because the 15-year AWS data
set used here is not sufficiently long to identify new
phenomena in this highly variable environment. Traditional
analysis techniques do not always perform well under these
conditions. The analyses here provide evidence for
expected links to ENSO [e.g., Bromwich et al., 2004;
Bromwich and Rogers, 2000; Savage et al., 1988], but
the work here was not optimized for study of ENSO
connections because of the use of annual averages here
versus the typical austral spring/summer ENSO signal that
crosses the calendar boundary. Further work at subannual
resolution, particularly the existing seasonal and monthly
AWS anomalies, is planned on this topic. Despite the lack
of major new insights here, noteworthy temperature anomalies, summer warming trends and links to lower latitude
climate phenomena identified in these records will motivate
further research. While it is difficult to say that all of these
results are simply not possible without these techniques, the
ANN-based methodology proved powerful and convenient
in conducting the research. Moreover, relative to competing, satellite-based techniques, such as passive microwavebased [Shuman and Stearns, 2001] or infrared-based
[Comiso, 2000] reconstructions, our methodology has fewer
24 of 28
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
Figure 16. Seasonal averages of ERA-15 sea level pressure: (a) spring 1980, (b) spring 1988, (c) spring
1986, and (d) fall 1982. Spring 1980 and 1988 had strong positive pressure anomalies at all six AWS
sites, while spring 1986 and fall 1982 had strong negative anomalies. The northern boundary in all plots
is 50S. Values are in millibars with a 4 mbar contour interval and every other contour labeled. Shading is
omitted for the range 984– 1000 mbar to emphasize extreme values. Latitude and longitude (see Figure 1)
are omitted for clarity.
25 of 28
D04103
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
Figure 17. Length of (a) winter and (b) summer seasons based on temperatures. Note change in axes
between winter and summer Figures.
26 of 28
D04103
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
D04103
Figure 17. (continued)
assumptions (e.g., about emissivity relationships) and
restrictions (e.g., only cloud-free days). It is also applicable
to near-surface pressure and, potentially, any other available
AWS measurement.
[45] Acknowledgments. This research was supported by the Office
of Polar Programs of the National Science Foundation through grants OPP
94-18622, OPP 95-26374, OPP 96-14927, and OPP 00-87380 to R. B.
Alley. We are also grateful to the Antarctic Meteorological Research Center,
University of Wisconsin, for their archive of Antarctic AWS data.
References
Breckenridge, C. J., U. Radok, C. R. Stearns, and D. H. Bromwich (1993),
Katabatic winds along the Transantarctic Mountains, in Antarctic Meteorology and Climatology: Studies Based on Automatic Weather Stations,
Antarct. Res. Ser., vol. 61, edited by D. H. Bromwich and C. R. Stearns,
pp. 69 – 92, AGU, Washington, D. C.
Bromwich, D. H., and J. J. Cassano (2001), Meeting summary: Antarctic
Weather Forecasting Workshop, Bull. Am. Meteorol. Soc., 82(7), 1409 –
1414.
Bromwich, D. H., and A. N. Rogers (2000), The El Niño—Southern
Oscillation modulation of West Antarctic precipitation, in The West
Antarctic Ice Sheet: Behavior and Environment, Antarct. Res. Ser.,
vol. 77, edited by R. B. Alley and R. Bindschadler, pp. 91 – 103, AGU,
Washington, D. C.
Bromwich, D. H., and C. R. Stearns (Eds.) (1993), Antarctic Meteorology
and Climatology: Studies Based on Automatic Weather Stations, Antarct.
Res. Ser., vol. 61, 207 pp., AGU, Washington, D. C.
Bromwich, D. H., F. M. Robasky, R. I. Cullather, and M. L. Vanwoert
(1995), The atmospheric hydrologic cycle over the Southern Ocean and
Antarctica from operational numerical analyses, Mon. Weather Rev.,
123(12), 3518 – 3538.
Bromwich, D. H., A. N. Rogers, P. Kållberg, R. I. Cullather, J. W. C. White,
and K. J. Kreutz (2000), ECMWF analyses and reanalyses depiction of
ENSO signal in Antarctic precipitation, J. Clim., 13, 1406 – 1420.
Bromwich, D. H., A. J. Monaghan, and Z. Guo (2004), Modeling the
ENSO modulation of Antarctic climate in the late 1990s with Polar
MM5, J. Clim., 17, 109 – 132.
Cassano, J. J., J. E. Box, D. H. Bromwich, L. Li, and K. Steffen (2001),
Evaluation of Polar MM5 simulations of Greenland’s atmospheric circulation, J. Geophys. Res., 106(D24), 3867 – 3889.
Cavazos, T. (1999), Large-scale circulation anomalies conducive to extreme
events and simulation of daily rainfall in northeastern Mexico and southeastern Texas, J. Clim., 12, 1506 – 1523.
Cavazos, T., A. C. Comrie, and D. M. Liverman (2002), Intraseasonal
variability associated with wet monsoons in southeast Arizona, J. Clim.,
15, 2477 – 2490.
Comiso, J. C. (2000), Variability and trends in Antarctic surface temperatures from in situ and satellite infrared measurements, J. Clim., 13,
1674 – 1696.
Crane, R. G., and B. C. Hewitson (1998), Doubled CO2 precipitation
changes for the Susquehanna basin: Down-scaling from the GENESIS
general circulation model., Int. J. Climatol., 18, 65 – 76.
Cullather, R. I., D. H. Bromwich, and R. W. Grumbine (1997), Validation
of operational numerical analyses in Antarctic latitudes, J. Geophys. Res.,
102(D12), 13,761 – 13,784.
Das, S. B., R. B. Alley, D. B. Reusch, and C. A. Shuman (2002), Temperature variability at Siple Dome, West Antarctica, derived from
ECMWF re-analyses, SSM/I and SMMR brightness temperatures and
AWS records, Ann. Glaciol., 34, 106 – 112.
Demuth, H., and M. Beale (2000), Neural Network Toolbox, 844 pp., Mathworks, Inc., Natick, Mass.
27 of 28
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
Gardner, M. W., and S. R. Dorling (1998), Artificial neural networks (the
multilayer perceptron): A review of applications in the atmospheric
sciences, Atmos. Environ., 32(14 – 15), 2627 – 2636.
Gibson, J. K., P. Kållberg, S. Uppala, A. Hernandez, A. Nomura, and
E. Serrano (1999), ERA-15 description, ECMWF Reanal. Rep. Ser., vol.
1, 84 pp., Eur. Cent. for Medium-Range Weather Forecasts, Reading, UK.
Hastenrath, S., and L. Greischar (1993), Further work on the prediction of
northeast Brazil rainfall anomalies, J. Clim., 6, 743 – 758.
Hastenrath, S., L. Greischar, and J. Van Heerden (1995), Prediction of the
summer rainfall over South Africa, J. Clim., 8, 1511 – 1518.
Haykin, S. S. (1999), Neural Networks: A Comprehensive Foundation, 842
pp., Prentice-Hall, Old Tappan, N. J.
Hewitson, B. C., and R. G. Crane (Eds.) (1994a), Neural Nets: Applications
in Geography, Kluwer Acad., Norwell, Mass.
Hewitson, B. C., and R. G. Crane (1994b), Precipitation controls in southern Mexico, in Neural Nets: Applications in Geography, edited by B. C.
Hewitson and R. G. Crane, pp. 121 – 143, Kluwer Acad., Norwell, Mass.
Holmes, R. E., C. R. Stearns, G. A. Weidner, and L. M. Keller (2000),
Utilization of automatic weather station for forecasting high wind speeds
at Pegasus Runway, Antarctica, Weather Forecast., 15(2), 137 – 151.
Middleton, G. V. (2000), Data Analysis in the Earth Sciences Using
Matlab, 260 pp., Prentice-Hall, Old Tappan, N. J.
Reusch, D. B. (2003), Nonlinear paleoclimatology: Reconstructions in West
Antarctica, Ph.D. thesis, 223 pp., Pa. State Univ., Univ. Park.
Reusch, D. B., and R. B. Alley (2002), Automatic weather stations and
artificial neural networks: Improving the instrumental record in West
Antarctica, Mon. Weather Rev., 130(12), 3037 – 3053.
D04103
Savage, M. L., C. R. Stearns, and G. A. Weidner (1988), The Southern
Oscillation signal in Antarctica, in Second Conference on Polar Meteorology and Oceanography, pp. 141 – 144, Am. Meteorol. Soc., Madison,
Wisc.
Shuman, C. A., and C. R. Stearns (2001), Decadal-length composite inland
West Antarctic temperature records, J. Clim., 14, 1977 – 1988.
Stearns, C. R., L. M. Keller, G. A. Weidner, and M. Sievers (1993),
Monthly mean climatic data for Antarctic automatic weather stations,
in Antarctic Meteorology and Climatology: Studies Based on Automatic
Weather Stations, Antarct. Res. Ser., vol. 61, pp. 1 – 22, AGU, Washington, D. C.
Turner, J., et al. (2004), The SCAR READER project: Towards a highquality data base of mean Antarctic meteorological observations, J. Clim.,
in press.
van Loon, H. (1967), The half-yearly oscillations in middle and high southern latitudes and the coreless winter, J. Atmos. Sci., 24(5), 472 – 486.
Wexler, H. (1958), The ‘‘kernlose’’ winter in Antarctica, Geophysica, 6(3 –
4), 577 – 595.
Wolff, E. W., J. S. Hall, R. Mulvaney, E. C. Pasteur, D. Wagenbach, and
M. Legrand (1998), Relationship between chemistry of air, fresh snow
and firn cores for aerosol species in coastal Antarctica, J. Geophys.
Res., 103(D9), 11,057 – 11,070.
R. B. Alley and D. B Reusch, Department of Geosciences, Pennsylvania
State University, 517 Deike Building, University Park, PA 16802-2712,
USA. ([email protected])
28 of 28
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
Figure 5. Daily average records from AWS sites of this study: (a) 2-m temperature and (b) 2-m
pressure. Observations are shown as a blue line, and ANN predictions are shown as a red line.
Confidence intervals derived from RMS errors appear as a light gray envelope around predictions.
10 of 28
D04103
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
Figure 5. (continued)
11 of 28
D04103
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
Figure 8. Monthly mean temperatures (C) in a month-by-year format, centered on austral winter
(January to December, x axis). Years on y axis refer to January. Values are interpolated in x and y in order
to improve the rendering; thus the figure should not be taken absolutely literally at the smaller scales. The
center of the color bar has been set to white to highlight values above/below the mean.
16 of 28
D04103
D04103
REUSCH AND ALLEY: WEST ANTARCTIC AWS CLIMATOLOGY
Figure 12. As in Figure 8 but for monthly mean pressure.
21 of 28
D04103