automatic weather stations and artificial neural networks
TRANSCRIPT
Automatic Weather Stations and Artificial Neural Networks:Improving the Instrumental Record in West Antarctica
David B. Reusch1 and Richard B. Alley
Department of Geosciences and EMS Environment Institute
The Pennsylvania State University
University Park, PA 16802 USA
1. Corresponding author, e-mail [email protected]
Submitted January 8, 2002
1
Abstract
Automatic weather stations (AWS) currently provide the only year-round, continuous
direct measurements of near-surface weather on the West Antarctic ice sheet away from
the coastal manned stations. Improved interpretation of the ever-growing body of ice-
core-based paleoclimate records from this region requires a deeper understanding of
Antarctic meteorology. As the spatial coverage of the AWS network has expanded year
to year, so has our meteorological database. Unfortunately, many of the records are
relatively short (less than 10 years) and/or incomplete (to varying degrees) due to the
vagaries of the harsh environment. Climate downscaling work in temperate latitudes
suggests that it is possible to use GCM-scale meteorological data sets (e.g., ECMWF
reanalysis products) to address these problems in the AWS record and create a uniform
and complete database of West Antarctic surface meteorology (at AWS sites). Such
records are highly relevant to the improved interpretation of the expanding library of
snow-pit and ice-core data sets.
Artificial neural network (ANN) techniques are used to predict AWS surface data
(temperature, pressure) using large-scale features of the atmosphere (e.g., 500 mb
geopotential height) from a region around the AWS. ANNs are trained with a calendar
year of observed AWS data (possibly incomplete) and corresponding GCM-scale data.
This methodology is sufficient both for high-quality predictions within the training set
and for predictions outside the training set that are at least comparable to the state-of-the-
art. For example, our results for temperature prediction are approximately equal to those
from a satellite-based methodology but with no exposure to problems from surface melt
events or sensor changes. Similarly, the significant biases seen in ECMWF surface
2
temperatures are absent from our predictions resulting in an RMS error half as large with
respect to the original AWS observations.
These results support high confidence in the ANN-based predictions from the GCM-scale
data for periods when AWS data are unavailable, e.g., before installation. ANNs thus
provide a means to expand our surface meteorological records significantly in West
Antarctica.
3
1. Introduction
To advance our knowledge of paleoclimate, we must improve our calibration of the ice
core-based proxies to the modern climate. This will improve our interpretive skill and
deepen our confidence in climate reconstructions. Because the climate that makes an ice
sheet a good recorder of climate also makes it inhospitable for humans and their weather
instruments, meteorological records from these regions are sparse and suffer greatly in
comparison to more temperate regions. Yet research in the temperate world has
suggested a new solution to this problem of short, interrupted, polar meteorological
records: artificial neural networks (ANNs). Similar to traditional climate downscaling
(e.g., Crane and Hewitson 1998), our ANN-based approach uses GCM-scale data to
predict surface meteorology based on the available surface record. But unlike most
climate downscaling work, the surface data from polar ice sheets are very limited.
Automatic weather stations (AWS) currently provide the only year-round, direct
measurements of weather away from the coast in West Antarctica (Figure 1). As the
spatial coverage of the network has expanded year to year, so has our meteorological
database, thus adding to our calibration data. Unfortunately, many of the records are
relatively short (less than 10 years) and/or incomplete (to varying degrees) due to the
vagaries of the harsh environment. Presuming that current AWS remain active, the
records will lengthen over time and eventually solve the shortness-of-record problem.
Equipment problems may also decline as improved instruments are deployed and existing
components upgraded. Nonetheless, for progress to occur in the near-term these
problems need to be addressed. Our ANN-based approach provides a means to both fill
gaps from instrument failures (and thereby improve the overall record quality) and to
4
extend records into time periods prior to AWS installation and after station
relocation/removal. In particular, we have used our ANN-based methodology to generate
complete records of pressure and temperature for Ferrell AWS (77.91¡ S, 170.82¡ E,
Figure 1) for the period 1979-1993. The new records are a merger of AWS observations
and ANN predictions for periods when observations were unavailable.
Our methodology is based on artificial neural networks (ANNs, Figure 2). An ANN is
composed of a large and highly-connected network of simple processing nodes organized
into layers and loosely modeled after neurons in the nervous system (e.g., Haykin 1999).
Nodes have multiple, weighted inputs and a single output (Figure 3). The weighted
inputs are combined and passed through a non-linear, often sigmoidal, activation function
to produce the output value. Multilayer feed-forward networks, the basic architecture of
our work, divide nodes into input, output, and hidden layers ( hidden because it is
internal to the network). In practice, feed-forward ANNs typically use three layers and
anywhere from just a few to hundreds of nodes per layer. Our methodology trains an
ANN with pairs of AWS observations and corresponding ECMWF variables from a
calendar year. The trained ANN is then used to predict missing AWS observations from
available ECMWF data, e.g., into years before the AWS was operational.
In Section 2 we describe the data used in our ANN-based prediction system. Further
details on the ANN architectures and training methods used are given in Section 3.
Section 4 presents analyses of our results and the new synthesized temperature and
pressure records for AWS Ferrell. Section 5 compares the ANN-based results to a
satellite-based temperature prediction technique and to ECMWF surface data.
5
2. Data
AWS Data
The main source of direct meteorological data in West Antarctica is the network of AWS
maintained by the University of Wisconsin-Madison since 1980 (Lazzara 2000). 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 main instrument cluster is nominally within 3 m
above the snow surface. This distance changes with snow accumulation and removal.
Pressure is calibrated to –0.2 hPa with a resolution of approximately 0.05 hPa.
Temperature accuracy is 0.25-0.5 ¡C with lowest accuracy at -70 ¡C, i.e., accuracy
decreases with decreasing temperature (M. Lazzara, pers. comm.). The data used here
are from the three hourly quality-controlled data sets available at the University of
Wisconsin-Madison FTP site (ice.ssec.wisc.edu). A subset of these data (for 0, 6, 12 and
18 UTC) is used to match ECMWF time-steps (see below).
Ferrell AWS was installed in December 1980 on the Ross Ice Shelf (77.91¡ S, 170.82¡
E), approximately 100 km east of McMurdo Station (Figure 1). Ferrell was selected here
because it has a longer and more continuous record (Figure 4) than most AWS, which are
generally more remote and harder to service. The 18 year (1981-1998) average
availability for temperature and pressure data, at three hourly resolution, is approximately
88%. Nine years exceed 95% availability while six years range from 65% to 75%. The
largest gaps in the study period (1979-1993), presumably from long-term equipment
problems, occur in the late austral winter/spring during 1983-1985 and 1991-1992
(Figure 4).
6
ECMWF Data
The ECMWF 15-year reanalysis data product (ERA-15) provided GCM-scale
meteorological data for the period 1979-1993 (ECMWF 2000). 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. Data are available at 0, 6, 12 and 18
UTC. A subset of the available variables (Table 1) and grid points was used at each time
step. Each grid point included all selected variables.
ERA-15 variables and grid selection
Table 1 summarizes ERA-15 variables used to predict AWS pressure and temperature.
Briefly, these variables were chosen because of their physical relationship to the
quantities being predicted. An exception to this guideline is the Julian decimal date used
in predicting temperature. This was added as a proxy for the strong annual signal seen in
temperature. The pressure levels selected represent the lower atmosphere over the station
and capture a substantial fraction of the regional circulation. ECMWF surface data have
not been used as a compromise between local and general predictive skill and to test the
utility of upper air data as a predictor for surface meteorology. This also allows us to use
the ECMWF surface data as a reference for predictive skill.
Several different configurations of grid points have been tried. The goal was to select a
subset of the lower atmosphere in the AWS region that is well-related to the surface
meteorology at the station itself and thus supports the predictive skill of the selected
predictor variables. Finding the best group of grid points is an exponentially hard
problem, which we have not attempted to solve. Instead, the focus has been on adjacent
7
points plus points from the corners of a square area centered approximately on the station
(Figure 5). Ferrell is located fortuitously close to an ECMWF grid point. Testing
showed that factors other than the grid point configuration have a substantially larger
influence on performance.
ERA-15 validity
Potential problems have been noted with the ECMWF (re)analysis data over Antarctica,
stemming in part from the flawed surface elevations used in these models (Genthon and
Braun 1995). Elevation errors exceeding 1000 m exist in some areas of Queen Maud
Land and the Antarctic Peninsula (e.g., Figure 3, Genthon and Braun 1995). Topography
in West Antarctica is generally much better but errors from outside our study area will
still have an influence on the reanalysis data (for example, an elevation error for Vostok
station has broad effects on geopotential heights). Evaluations of several operational
products (e.g., Bromwich et al. 1995; Bromwich et al. 2000; Cullather et al. 1998) and
discussions with experienced polar meteorologists (D. Bromwich, J. Turner, pers. comm.)
suggest that the ECMWF analyses are the best data sets currently available for Antarctica
(see also Bromwich et al. 1998). This is expected to remain true until such time as the
currently-in-progress ECMWF 40-year reanalysis is readily available (ECMWF 2001).
3. Methods
At the simplest level, artificial neural networks (ANNs) are a computer-based problem
solving tool inspired by the original, biological neural network — the brain. Because of
their ability to generate non-linear mappings during training, ANNs are particularly well-
suited to complex, real-world problems such as understanding climate (Elsner and Tsonis
1992; Tarassenko 1998). Meteorological examples include an improved understanding
8
of controls on precipitation in southern Mexico (Hewitson and Crane 1994), prediction of
summer rainfall over South Africa (Hastenrath et al. 1995) and northeast Brazil
(Hastenrath and Greischar 1993), and extreme event analysis in the Texas/Mexico border
region (Cavazos 1999). Our ANNs were implemented with the MATLAB Neural
Network Toolbox (Demuth and Beale 2000; Haykin 1999). Separate ANNs are currently
used for each AWS variable due to the different physical controls involved.
ANN Architectures
Three ANN types were used, all variants of the basic multilayer feed-forward ANN
(Figure 2). All share the same general form of processing node (Figure 3) but use
differing connectivity and activation functions. The multilayer feed-forward (FF) ANN
was selected because of its widespread use in predictive tasks and to follow previous
work with climate downscaling in the literature (e.g., Cavazos 1999). The three variants,
radial basis, general regression (GRNN) and Elman, offer different approaches to the
prediction problem. FF ANNs consist of a large number of highly interconnected, simple
processing nodes (a.k.a. neurons) organized into at least three layers (Figure 2). The
input layer serves to receive input data, with one node for each input variable. The output
layer receives intermediate results from the hidden layer and translates them to the
desired output format. The intermediate, or hidden, layer nodes take inputs from the
preceding layer, usually nodes of the input layer, and pass output to the subsequent layer,
usually nodes of the output layer. The number of hidden nodes is both problem- and
architecture-dependent and is a significant factor in how well the ANN works. Too many
nodes can lead to overfitting while too few will result in the network not learning the
problem effectively. Processing within each node consists of three steps: 1) each input is
9
multiplied by an input-dependent weight, 2) the weighted values and a node-dependent
bias (possibly zero) are summed, and 3) the result is passed to a non-linear, often
sigmoidal (e.g., tanh), activation function. The output of the activation function
determines the output of the node.
Elman networks add to the FF ANN a feedback from the hidden layer output to the
hidden layer input. Adding this recurrent connection allows this type of ANN to detect
(and generate) time-varying patterns (Demuth and Beale 2000). Our experience suggests
that this feature was of no particular benefit to our problem, though this is likely due to
our algorithm for selecting the training records. With our algorithm, the Elman network
appears to behave like a slightly improved FF ANN.
Radial basis ANNs make a number of changes to the FF ANN design. First, only one
hidden layer is ever used. Second, multiplication and summation is replaced by
calculation of the vector distance between an input vector and the weight vector
associated with each hidden layer node. This yields a vector of distances between the
input pattern and each node s weight vector. The distance vector and the bias are then
multiplied element-wise to adjust the sensitivity of each node. Third, the sigmoid
activation function is replaced by a radial basis function of the form exp(-n2), where n is
the result of the preceding computational steps. The net result of these changes is that a
node will only activate for input patterns closely matching its weight vector (Demuth and
Beale 2000). This means that radial basis nodes only respond to relatively small areas of
the input space, unlike the sigmoidal nodes of FF ANNs. Because of these differences,
radial basis ANNs typically require more hidden layer nodes than FF ANNs, but they
can, in theory, be trained more quickly (Demuth and Beale 2000).
10
General regression neural networks are a variant of radial basis ANNs and are often used
for function approximation (Demuth and Beale 2000). As before, the hidden layer uses
radial basis functions but with one node for each input vector (this is also sometimes
done for standard radial basis ANNs). The GRNN also modifies the computations in
each output layer node. First, the node weights are fixed during training to be the target
vectors associated with the input vectors. Second, when the ANN is processing input, the
output nodes first compute the dot product of the hidden layer output vector and the
output node weight vector. This value is normalized by the sum of the elements of the
hidden layer output vector before being passed to the linear activation function to
produce the final output value. Thus an input closely matching an input/target pair used
in training will first produce a hidden layer node with an output close to 1. The output
layer then translates that node to the closest original target from training. Outputs for
input values not seen in training depend on the sensitivity of the radial basis nodes.
All results presented here are derived from the best-performing Elman ANNs for
temperature and pressure. The general regression and radial basis ANNs produced
comparable, but slightly poorer performance. That the three techniques performed
comparably supports the suitability of the ANN approach.
ANN Training and Testing
Our methodology revolves around finding an ANN best suited to predicting an AWS
variable using some set of ECMWF variables as input. This task can be broken down
into three nested/overlapping subtasks: training individual ANNs, creating ensembles of
ANNs with the same inputs, and searching for the best set of input predictors and non-
data dependent ANN parameters.
11
ANN Training
FF ANNs need to be taught to produce the desired outputs (AWS observations) from the
inputs (ECMWF data) before they can be used for predictions, a task done iteratively in
three main phases: training, testing and validation (a step dependent on the ANN
architecture). The training phase adjusts the connection weights using an optimization
function that reduces the error in the network s results. Training records are selected
randomly from the set of input observations (covering one calendar year) and represented
between 30% and 70% of the input records. The training error is calculated by
comparing the network s output prediction to the AWS observations for all input/target
pairs. Weights in each layer are then adjusted with a backpropagation algorithm using
the cumulative error from one pass through the complete training set. Testing uses a
second subset (typically 20%) of the input data to evaluate training performance at the
end of each training iteration. Validation is used to avoid overfitting the training data and
tests the network with data distinct from the training and testing samples. Depending on
the architecture being trained, validation used 10% of the input or was done outside the
training/testing cycle with observations from different calendar years. The cycle then
repeats until the desired output is achieved or the error cannot be further reduced or
begins to go up significantly. Details of the training process vary between architectures.
Ensembles
Wrapped metaphorically around individual ANN training is a loop for training from
different initial conditions and training records. The extraordinary number of parameters
involved in ANNs (each weight, bias and input combines multiplicatively) leads to a
highly complex, multidimensional error surface with numerous local minima. Because
12
of this, it is very important to train multiple versions of the same configuration using
different initial weights and/or training records. We achieve this by running a large
number of iterations (typically 50) of the same ANN configuration. Each instance of the
ANN starts from different randomly initialized weights and is trained with a different
randomly selected set of input data. The top 10 networks (by RMS error) were saved for
further testing. While it has not yet been implemented, some performance improvement
might be gained by stacking the results from, for example, the top 5 best ANNs from the
overall best performing configuration.
Experimental Design
Selection of a best ANN involves numerous dimensions of possible parameters. In the
physical domain, a variety of predictor variables are available (e.g., geopotential height,
thickness) as well as multiple pressure levels. Selection of appropriate grid points adds a
second physical dimension, though results have not been particularly sensitive to our
choices. We have explored many of these dimensions by wrapping an experimental
design loop around the above training/testing process for individual ANNs. Using this
logical loop, we were able to identify the most useful pressure levels and variables.
Optimal grid point selection would also happen in this loop. With the exponential nature
of that task, we have opted to work with a useful set of points rather than the optimal
set. There are also a number of logical dials that can be adjusted to optimize ANN
performance, such as learning rate and momentum, and testing of these variables was
done at this level. Table 1 summarizes the ECMWF variables used for the best ANNs.
13
4. Results
The nature of our current ANN training methodology, where ANNs are trained with one
calendar year of AWS observations, leads to excellent performance within the training
year but a diminished skill for all other years. This approach is used to demonstrate the
extreme case of only one year s worth of AWS data being available for ANN training, a
situation likely to be true for the most recently installed AWSs. Few of the existing AWS
have the record length available at Ferrell so it is useful to test the methodology under
worst-case conditions. We also initially believed that one year of training data would be
sufficient to obtain acceptable predictive skill. This is not entirely false, but it is also
clear that performance would likely improve by taking advantage of more training data if
available.
Statistics
Figure 6 graphically summarizes surface temperature prediction results for a training year
(1987) and an arbitrary non-training year (1983). Figure 7 does the same for surface
pressure. In each case, the training year results are noticeably better than the non-training
year results.
Prediction results for all years are summarized in Figure 8. Although most non-training
years have lower performance than the training year results, there are some non-training
years that do nearly as well predicting with the training year ANN. Also shown in Figure
8 are results from training with a different calendar year. Again, the training year has the
best performance and other years do worse. We have tried a number of different training
years (1982, 1987, 1990 and 1993) but have not seen any distinct patterns of performance
14
in the other years or any particular benefit to any given year. In short, all training years
seem to give roughly the same results for non-training years.
Table 2 summarizes seasonal statistics for the ANN predictions. Results were analyzed
on a seasonal basis to determine if the ANN performance had any relationship to the time
of year. The statistics show a small, possibly negligible effect for pressure with spring
and summer having slightly better results than fall and winter. The results for
temperature appear more compelling with an apparently significant difference between
higher summer and lower winter predictive skill. Spring and fall are nearly identical to
the average RMS error. Thus there may be some relationship to season for temperature
but it is not year-round.
New Records
After completing the ANN training process, the best ANNs were used to synthesize 15-
year records of pressure and temperature for this AWS. As with Shuman and Stearns
(2001), the final records are a merger of AWS observations and ANN predictions for
those periods where observations are unavailable. A measure of uncertainty for the
predictions was generated by using the ANN to predict available AWS observations for
each year and calculating the RMS error of the predictions. For those years where no
observations are available (e.g., before AWS installation), the average RMS error was
used. This provides the basis for the error bars in Figure 9.
5. Discussion
To further assess the quality of our methodology, we would like to compare our results
with the best available results from similar attempts to improve the AWS record.
15
Unfortunately, very little has been published on this subject in Antarctica. One
alternative to our ANN-based technique uses satellite passive-microwave brightness
temperatures (Shuman and Stearns 2001). In lieu of other alternate techniques, it is also
reasonable to compare our performance to available model results, such as the ECMWF
surface data.
Comparison to a satellite-based technique
The recent work by Shuman and Stearns (2001, hereafter SS) used satellite passive
microwave brightness temperatures and approximate surface emissivity to reconstruct
surface temperatures at a number of AWS in West Antarctica. In the satellite-based
methodology, three-hourly AWS observations were first averaged to daily values before
comparison with the daily passive-microwave brightness temperatures. Our technique
produces a calculated surface temperature for all available six-hourly AWS observations
(for ECMWF time steps at 0, 6 12 and 18 UTC) thus yielding up to four predictions each
day (for those days with no missing six-hourly AWS observations). Thus both our
observed and calculated daily means are based on up to four six-hourly values whereas
only one daily calculated value is available from the SS technique.
Error analyses documented in SS include comparisons of calculated and observed surface
temperatures on a daily and annual basis. Figure 10 shows our mean daily surface
temperatures. There is a hint of improved predictive accuracy at higher temperatures but
this does not hold up to closer examination (e.g., a probability density function did not
show a strong difference for higher temperature predictions). There are otherwise no
distinct artifacts such as the curvature (from some calculated temperatures being too low
in spring and fall or too high in summer and winter) seen in some SS results (SS, Figure
16
10). Table 3 summarizes statistics based on differences between the calculated and
observed daily surface temperatures. The values in the first three columns reproduce SS
Table 4. The column headed Training shows results from the best network trained on
1987 observations. The remaining columns summarize performance of the same network
on all other years. ANN performance is clearly comparable to the SS methodology for
the training year. For the remaining years, the mean error in our results is larger (0.56 ¡C
versus 0.14 ¡C) but the standard deviation (σn-1) is nearly identical (5.48 ¡C versus 5.52
¡C). As the training year results demonstrate, the errors from the ANN-based
methodology could be greatly reduced by using one ANN per year at the expense of
greater complexity. Improvement may also be possible by refining the method used for
selection of training records.
The transfer function used in the SS methodology depends on a modeled emissivity to
convert passive-microwave brightness temperatures to surface air temperature. The
accuracy of this transfer function is thus a significant contributor to the overall accuracy
of the calculated temperature values. The transfer function is based on temporally
overlapping brightness temperatures and AWS observations that are used to generate a
modeled emissivity time series. Thereafter, surface temperature is estimated from the
brightness temperatures via the emissivity time series. Significant departures in
microwave brightness temperatures can arise due to melt events and associated liquid
water in the snowpack, and to the density contrast remaining when the liquid water
refreezes. This may lead to incorrect calculated surface air temperatures if the transfer
function does not adjust for the changed relationship between brightness temperature and
air temperature (e.g., SS, Figure 9). This type of error might be reduced by including
17
surface temperature data associated with melt events in the transfer function calibration
process. Suitable data may not always be available, however. Our methodology should
be immune to errors due to melt events since it does not rely on characteristics of the
snow surface. A site, such as Lettau, with observed temperatures near to above freezing
could be used to confirm this assumption (Ferrell observations are all below freezing).
As pointed out in SS, however, only merged records with substantial missing summer
temperature observations are likely to be susceptible to melt event related errors.
Annual averages of calculated and observed daily surface temperatures were also
analyzed in SS. Figure 11a compares annual means at Ferrell AWS for those years with
at least 340 days of observations; Figure 11b shows all years (1981-1993). Differences
between annual averages of calculated and observed values are all less than 1.5 ¡C. Table
4 summarizes the statistics of this comparison. The mean error in the ANN-based
methodology is directly comparable to the errors in the SS technique (their Table 5). Our
standard deviation (σn-1) is at the high end of the SS range but still reasonably low.
Adding in years with fewer than 340 days of observations (the lowest being 1992 with
~236 valid days) does not change the mean or standard deviation significantly.
The ANN-based technique compares well with the satellite-based approach. Our
approach is also immune to melt event-related problems, has minimal exposure to
changes in sensors, and is based on data (ECMWF) with no gaps. Furthermore, our
technique is also applicable to surface pressure.
18
Evaluation of ECMWF surface data
ECMWF surface data could also be used to fill the gaps in the AWS records and would
appear to be just as reasonable as any data from an empirical methodology. They also
provide an alternative benchmark since these data were not used in our ANN training.
The nearest ECMWF grid point to Ferrell AWS is at 77.5 ¡S, 170¡E, approximately 50
km away (Figure 5). A comparison of 2-meter temperatures from this grid point to the 2-
m temperatures observed at the AWS (Figure 12) reveals the flaws in the ECMWF data.
While the correlation (0.91) and standard deviation (5.2 ¡C) are similar to our ANN-
based results (Figure 10), the RMS error is approximately twice as large (10.3 ¡C versus
5.4 ¡C). There are also clear biases in the ECMWF data not present in our predictions.
Thus our ANN-based predictions of surface temperature are also superior to the ECMWF
model data.
The ECMWF data fare better for surface pressure (Figure 13) and are comparable to our
ANN-based predictions (Figure 14). Our RMS error (2.9 mbar versus 4.7 mbar) and
mean error (0.41 mbar versus 4.13 mbar) are better, suggesting a benefit, albeit possibly
slight, to our methodology for this variable.
6. Conclusion
This work has shown the utility of an ANN-based approach to predicting AWS
observations of near-surface temperature and pressure using variables derived from
GCM-scale numerical forecast models. With the current methodology, skill within the
training year is high while predictions outside the training year are of moderately lower
quality. This is not seen as a major issue since there are still alternative training methods
and approaches remaining to be explored.
19
The ANN-based technique also compares well with the satellite-based approach. Our
approach should be immune to melt event-related problems, has minimal exposure to
changes in sensors, and is based on data (ECMWF) with no gaps. Our results also do not
appear to be strongly seasonally biased, although there may be a minor seasonal
dependence for temperature. Furthermore, our technique is also applicable to surface
pressure. Lastly, we will be able to extend our methodology into the pre-satellite era
once the ECMWF 40-year reanalysis data sets become available.
Our results also compare well to the ECMWF surface data. These data are not used in
our methodology so independent comparisons can be made to the AWS observations.
Our temperature predictions have an RMS error approximately one-half that of the
ECMWF surface data without the biases present in the latter. This suggests that while the
upper air data may have similar imperfections, the ANN technique is not sensitive to
them. While this may be true in a sense, it is also possible that improvements in the
quality of the upper air data will require revisiting the ANN training process so that the
relationship to the AWS observations can be relearned.
By using one calendar year of training data, we have shown what can be expected from
applying this technique to AWS with short observational records. This should also be the
worst case for those AWS with longer records. Further research will explore using more
of the available record in training for those sites where this is an option, including Ferrell
AWS itself.
20
References
Bromwich, D. H., R. I. Cullather, et al., 1998: Antarctic precipitation and its contribution
to the global sea-level budget. Ann. Glaciol., 27, 220-226.
Bromwich, D. H., F. M. Robasky, et al., 1995: The atmospheric hydrologic cycle over the
Southern Ocean and Antarctica from operational numerical analyses. Mon. Wea.
Rev., 123, 3518-3539.
Bromwich, D. H., A. N. Rogers, et al., 2000: ECMWF Analyses and Reanalyses
Depiction of ENSO Signal in Antarctic Precipitation. J. Climate, 13, 1406-1420.
Cavazos, T., 1999: Large-scale circulation anomalies conducive to extreme events and
simulation of daily rainfall in northeastern Mexico and southeastern Texas. J.
Climate, 12, 1506-1523.
Crane, R. G. and B. C. Hewitson, 1998: Doubled CO2 precipitation changes for the
Susquehanna basin: Down-scaling from the GENESIS general circulation model.
Inter. J. Climatol., 18, 65-76.
Cullather, R. I., D. H. Bromwich, et al., 1998: Spatial and temporal variability of
Antarctic precipitation from atmospheric methods. J. Climate, 11, 334-367.
Demuth, H. and M. Beale, 2000: Neural Network Toolbox. Mathworks, Inc., 844 pp.
ECMWF, cited 2001: ERA-15. [Available online from
http://wms.ecmwf.int/research/era/Era-15.html.]
, cited 2001: ERA-40 Project Plan. [Available online from
http://wms.ecmwf.int/research/era/Project_plan.html.]
21
Elsner, J. B. and A. A. Tsonis, 1992: Nonlinear Prediction, Chaos, and Noise. Bull. Amer.
Meteor. Soc., 73, 49-60.
Genthon, C. and A. Braun, 1995: ECMWF Analyses and Predictions of the Surface
Climate of Greenland and Antarctica. J. Climate, 8, 2324-2332.
Hastenrath, S. and L. Greischar, 1993: Further Work on the Prediction of Northeast
Brazil Rainfall Anomalies. J. Climate, 6, 743-758.
Hastenrath, S., L. Greischar, et al., 1995: Prediction of the Summer Rainfall over South
Africa. J. Climate, 8, 1511-1518.
Haykin, S. S., 1999: Neural networks : a comprehensive foundation. Prentice Hall, 842
pp.
Hewitson, B. C. and R. G. Crane, 1994: Precipitation controls in southern Mexico.
Neural Nets: Applications in Geography, B. C. Hewitson and R. G. Crane, Eds.,
Kluwer Academic, 121-143.
Lazzara, M. A., cited 2000: Antarctic Automatic Weather Stations Web Site Home Page.
[Available online from http://uwamrc.ssec.wisc.edu/aws/.]
Shuman, C. A. and C. R. Stearns, 2001: Decadal-length composite inland West Antarctic
temperature records. J. Climate, 14, 1977-1988.
Tarassenko, L., 1998: A Guide to Neural Computing Applications. John Wiley & Sons,
Inc., 139 pp.
22
Tables
Table 1. Variables used to predict AWS near-surface observations of pressure and
temperature. Geopotential height (m), wind speed (m/s) and wind direction (¡) are from
ECMWF datasets. Thickness (m) and temperature advection (¡C/km) are derived from
ECMWF data. Julian decimal date is simply the day of the year divided by the total days
in the year.
Table 2. Seasonal statistics for temperature and pressure prediction errors (1987 training
year). Seasons are defined for southern hemisphere as Dec-Feb (summer), Mar-May
(fall), Jun-Aug (winter), and Sep-Nov (spring). n refers to the number of data points in
each time period. r refers to a simple linear correlation.
Table 3. Prediction error (difference between calculated and observed) statistics for daily
mean temperatures (¡C). Values in first three data columns from Table 4 of Shuman and
Stearns (2001). The mean is the average of the error for their four AWS (Byrd, Lettau,
Lynn and Siple). The minimum and maximum are the extrema of the published values.
Remaining columns are from our work with AWS Ferrell. Training values are from the
ANN training year (1987) using the best ANN. Values in the remaining columns are for
all other years between 1981 and 1993. The 1987 ANN was used to calculate
temperatures for these years. Each year was predicted individually with statistics
calculated by year.
Table 4. Statistics for differences between calculated and observed daily mean
temperatures (¡C) on an annual average basis. Values in first two columns from Table 5
of Shuman and Stearns (2001) and are based on results from four AWS (Byrd, Lettau,
23
Lynn and Siple). Remaining values based on AWS Ferrell. Only years with at least 340
days of observations were included in data columns one to three (seven years for our
results). All 13 years were included in data column four (1979 and 1980 were omitted
because they had no observations).
24
Figures
Figure 1. Location map for Antarctic automatic weather stations (AWS) described in
text. Ferrell AWS is the subject of this study. Siple, Byrd, Lettau and Lynn AWS were
studied in Shuman and Stearns (2001). Remaining labels for reference. Ferrell AWS
was installed on the Ross Ice Shelf (77.91¡ S, 170.82¡ E) in December 1980 (no 1980
observations have been used in this work).
Figure 2. Generalized multi-layer feed-forward artificial neural network (ANN).
Figure 3. Sample artificial neural network processing node with three inputs, a sigmoidal
activation function and no bias.
Figure 4. Ferrell AWS 6-hourly observations of (a) 2-m pressure and (b) 2-m air
temperature, both for 1981-1993. Data extracted from the three hourly quality-controlled
AWS archive data sets available at the University of Wisconsin-Madison FTP site
(ice.ssec.wisc.edu). Time steps 0, 6, 12 and 18 UTC selected to match ECMWF data.
Figure 5. ECMWF grid points around Ferrell AWS (central black square). Light lines
show ECMWF 2.5¡ x 2.5¡ horizontal grid. Black circles show sample grid point
locations used for training and prediction. Base AVHRR image by Matt Lazzara,
University of Wisconsin
Figure 6. Temperature prediction results for a training year (1987: a, b) and a non-
training year (1983: c, d). In the scatter plots (a, c), the thin solid line from lower left to
upper right is the ideal 1:1 line where all points would fall with perfect predictive skill.
The thicker solid line is a linear regression through the data points (the equation of this
25
line is shown in the legend box). Thin dashed lines are offset one RMS error from the
ideal 1:1 line to help show spread in the error. Standard statistics related to the difference
between observed and predicted are shown in the upper left corner. The absolute
prediction error (predicted - observed) is summarized in the error distribution plots (b, d).
The thin sloping line represents a normal distribution. Offsets from this line are offsets
from a normal distribution in the error. The vertical line is placed at the RMS error.
Figure 7. As in Figure 6 but for pressure predictions.
Figure 8. Summary of RMS errors for all years: (a) best temperature ANN trained with
1987 observations, (b) best pressure ANN also trained with 1987 observations, (c) best
pressure ANN trained with 1982 observations. Dashed horizontal line is the mean RMS
for all years excluding the training year.
Figure 9. Reconstructed six-hourly (a) surface pressure and (b) temperature at Ferrell
AWS for 1979-1993 (original observations as thin line, ANN-modeled values as points
with error bars). The ANN was trained with 1987 data. ECMWF data were used to fill
gaps and extend record back to 1979. Error bars on predictions are based on the RMS
error for the calendar year of the predictions (for 1981-1993) or the average RMS error
for 1981-1993 (for 1979-1980).
Figure 10. Daily mean calculated surface temperatures versus observations, 1981-1993.
Lines as in previous scatter plots.
26
Figure 11. Annual mean calculated surface temperatures versus observations: (a) only
years with at least 340 days of observations; (b) all years, 1981-1993. Lines as in
previous scatter plots.
Figure 12. ECMWF 2-m temperatures from grid point 77.5 ¡S, 170¡E compared to
observed 2-m temperatures at AWS Ferrell (77.91¡ S, 170.82¡ E). Both data sets are
daily averages of 6-hourly data. Lines as in previous scatter plots.
Figure 13. ECMWF surface pressures from grid point 77.5 ¡S, 170¡E compared to
observed surface pressures at AWS Ferrell (77.91¡ S, 170.82¡ E). Both data sets are
daily averages of 6-hourly data. Lines as in previous scatter plots.
Figure 14. Daily mean calculated surface pressures versus observations, 1981-1993.
Lines as in previous scatter plots.
27
Pressure Temperature
850 mb geopotential height 850 mb geopotential height
700 mb geopotential height 850 mb temperature advection
700-850 mb thickness 700-850 mb thickness
850 mb wind speed, direction Julian decimal date
28
Temperature Pressure
Period n RMSE (¡C) r n RMSE (mbar) r
All data 16281 6.4 0.85 16305 3.3 0.96
Winter 4252 7.7 0.51 4260 3.7 0.96
Spring 3424 6.6 0.78 3424 2.9 0.96
Summer 3935 4.6 0.79 3948 2.5 0.94
Fall 4670 6.3 0.67 4673 3.7 0.94
29
This Work
Shuman and Stearns Non-Training Years
Statistics Minimum Mean Maximum
Training
Year Mean σn-1 Minimum Maximum
Mean 0.094373 0.14198 0.19882 0.01 0.56 1.11 -1.95 1.78
σn-1 4.0027 5.5249 6.2450 3.0 5.48 0.44 4.8 6.1
30
Shuman and Stearns This Work
Statistics Minimum Maximum Best Years All Years
Mean 0.076 0.489 0.45 0.50
σn-1 0.209 1.001 1.3 1.1
1979
1980
1981
1982
1983
1984
940
960
980
1000
1020
Fer
rell
AW
S 6
−ho
urly
Pre
ssur
e 19
79−
1993
mbar
1984
1985
1986
1987
1988
1989
940
960
980
1000
1020
mbar
1989
1990
1991
1992
1993
1994
940
960
980
1000
1020
mbar
1979
1980
1981
1982
1983
1984
−60
−50
−40
−30
−20
−100
Fer
rell
AW
S 6
−ho
urly
Tem
pera
ture
197
9−19
93
deg C
1984
1985
1986
1987
1988
1989
−60
−50
−40
−30
−20
−100
deg C
1989
1990
1991
1992
1993
1994
−60
−50
−40
−30
−20
−100
deg C
−50 −40 −30 −20 −10 0 10
−50
−40
−30
−20
−10
0
10
Training Year (1987)
AWS Observation
AN
N P
redi
ctio
n
RMSE=4.1 deg Cr=0.94Mean=−0.07 deg CStdDev=4.1 deg C
a.
Ideal 1:1y=0.88x−3.02
0 5 10 15
0.0010.0030.01 0.02 0.05 0.10
0.25
0.50
0.75
0.90 0.95 0.98 0.99
0.9970.999
Absolute Error (deg C)
Pro
babi
lity
Training Year (1987)
RM
SE
=4.
1
b.
−60 −50 −40 −30 −20 −10 0
−60
−50
−40
−30
−20
−10
0
Non−training Year (1983)
AWS Observation
AN
N P
redi
ctio
n
RMSE=6.2 deg Cr=0.86Mean=0.78 deg CStdDev=6.1 deg C
c.
Ideal 1:1y=0.77x−4.79
0 5 10 15 20
0.0010.0030.01 0.02 0.05 0.10
0.25
0.50
0.75
0.90 0.95 0.98 0.99
0.9970.999
Absolute Error (deg C)
Pro
babi
lity
Non−training Year (1983)
RM
SE
=6.
2
d.
940 960 980 1000
940
950
960
970
980
990
1000
1010
Training Year (1987)
AWS Observation
AN
N P
redi
ctio
n
RMSE=2.3 mbarr=0.98Mean=−0.05 mbarStdDev=2.3 mbar
a.
Ideal 1:1y=0.95x+46.73
0 5 10 15
0.0010.0030.01 0.02 0.05 0.10
0.25
0.50
0.75
0.90 0.95 0.98 0.99
0.9970.999
Absolute Error (mbar)
Pro
babi
lity
Training Year (1987)
RM
SE
=2.
3
b.
960 980 1000 1020
950
960
970
980
990
1000
1010
1020
1030Non−training Year (1983)
AWS Observation
AN
N P
redi
ctio
n
RMSE=3.1 mbarr=0.96Mean=0.36 mbarStdDev=3.0 mbar
c.
Ideal 1:1y=0.94x+60.65
0 5 10 15
0.0010.0030.01 0.02 0.05 0.10
0.25
0.50
0.75
0.90 0.95 0.98 0.99
0.9970.999
Absolute Error (mbar)
Pro
babi
lity
Non−training Year (1983)
RM
SE
=3.
1
d.
0
2
4
6
8
1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993
Temperature - 1987 ANN
7.2
6.1 6.26.5
7.36.8
4.1
6 66.5
6.2
7.3 7.1
RM
S E
rror
(de
g C
)
Year
Training YearMean6.6
0
1
2
3
4
5
1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993
Pressure - 1987 ANN4.7
3.43.1
3.73.3
2.8
2.3
2.8 2.83
2.2
3.53.9
RM
S E
rror
(m
bar)
Year
Training YearMean
3.3
0
1
2
3
4
5
1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993
Pressure - 1982 ANN4.7
2.5
3.53.9 3.9
3.3 3.4 3.3 3.4 3.4
2.8
3.84.1
RM
S E
rror
(m
bar)
Year
Training Year Mean3.6
a.
b.
c.
1979
1980
1981
1982
1983
1984
940
960
980
1000
1020
Fer
rell
AW
S 6
−ho
urly
Pre
ssur
e 19
79−
1993
mbar
1984
1985
1986
1987
1988
1989
940
960
980
1000
1020
mbar
1989
1990
1991
1992
1993
1994
940
960
980
1000
1020
mbar
1979
1980
1981
1982
1983
1984
−60
−50
−40
−30
−20
−100
Fer
rell
AW
S 6
−ho
urly
Tem
pera
ture
197
9−19
93
deg C
1984
1985
1986
1987
1988
1989
−60
−50
−40
−30
−20
−100
deg C
1989
1990
1991
1992
1993
1994
−60
−50
−40
−30
−20
−100
deg C
−60
−50
−40
−30
−20
−10
0
−60
−50
−40
−30
−20
−100
Tem
pera
ture
198
1 −
199
3, D
aily
ave
rage
AW
S O
bser
vatio
n
ANN Prediction
RM
SE
=5.
4 de
g C
r=0.
89M
ean=
0.55
deg
CS
tdD
ev=
5.4
deg
C
Idea
l 1:1
y=0.
81x−
4.22
−27
−26
−25
−24
−23
−22
−27
−26
−25
−24
−23
−22
Ann
ual A
vera
ge T
empe
ratu
re 1
981
− 1
993
AW
S O
bser
vatio
n
ANN Prediction
RM
SE
=1.
3 de
g C
r=0.
59M
ean=
0.45
deg
CS
tdD
ev=
1.3
deg
C
Yea
rs w
ith a
t lea
st 9
3% d
ata
Idea
l 1:1
y=0.
57x−
10.2
9
−30
−28
−26
−24
−22
−20
−30
−28
−26
−24
−22
−20
Ann
ual A
vera
ge T
empe
ratu
re 1
981
− 1
993
AW
S O
bser
vatio
n
ANN Prediction
RM
SE
=1.
2 de
g C
r=0.
90M
ean=
0.50
deg
CS
tdD
ev=
1.1
deg
C
All
year
s
Idea
l 1:1
y=0.
75x−
5.79
−60
−50
−40
−30
−20
−10
0
−60
−50
−40
−30
−20
−100
EC
MW
F v
s A
WS
Dai
ly A
vera
ge T
empe
ratu
re 1
981
− 1
993
AW
S O
bser
vatio
n
ECMWF Surface Prediction
RM
SE
=10
.3 d
eg C
r=0.
91M
ean=
8.86
deg
CS
tdD
ev=
5.2
deg
C
Idea
l 1:1
y=0.
70x+
1.31
930
940
950
960
970
980
990
1000
1010
1020
1030
930
940
950
960
970
980
990
1000
1010
1020
1030
EC
MW
F v
s A
WS
Dai
ly A
vera
ge P
ress
ure
1981
− 1
993
AW
S O
bser
vatio
n
ECMWF Surface Prediction
RM
SE
=4.
7 m
bar
r=0.
98M
ean=
4.13
mba
rS
tdD
ev=
2.3
mba
r
Idea
l 1:1
y=0.
95x+
53.9
4
940
950
960
970
980
990
1000
1010
1020
1030
940
950
960
970
980
990
1000
1010
1020
1030
Pre
ssur
e 19
81 −
199
3, D
aily
ave
rage
AW
S O
bser
vatio
n
ANN Prediction
RM
SE
=2.
9 m
bar
r=0.
96M
ean=
0.41
mba
rS
tdD
ev=
2.9
mba
r
Idea
l 1:1
y=0.
95x+
50.9
9