Automatic weather stations (AWSs) 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 the meteorological database. Unfortunately, many of the records are relatively short (less than 10 yr) 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 datasets (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 datasets.
Artificial neural network (ANN) techniques are used to predict 6-hourly 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, the results presented herein 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 temperatures are absent from the predictions here, resulting in an rms error that is 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, for example, before installation. ANNs thus provide a means to expand the surface meteorological records significantly in West Antarctica.
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 temperature 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 (AWSs) currently provide the only year-round, direct measurements of weather away from the coast in West Antarctica (Fig. 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 yr) and/or incomplete (to varying degrees) due to the vagaries of the harsh environment. Presuming that current AWSs 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 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 the Ferrell AWS (77.91°S, 170.82°E; Fig. 1) for the period 1979–93. The new records are a merger of AWS observations and ANN predictions for periods when observations were unavailable.
Our approach follows that of a growing number of other authors (e.g., Calvo et al. 2000; Cavazos 1999, 2000; Trigo and Palutikof 1999) in using multilayer feed-forward ANNs (Fig. 2) to predict surface meteorology from gridded datasets of upper-atmospheric variables. Briefly, an ANN is composed of a (usually) 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 (Fig. 3). In practice, feed-forward ANNs typically use three layers and anywhere from just a few to hundreds of nodes per layer. ANNs have been used for downscaling of temperature (e.g., Huth 1999) and precipitation (e.g., Cavazos 1999, 2000) as well as in the analysis of other climate variables (e.g., Cannon and McKendry 1999; D'Odorico et al. 2000). We believe our work to be the first use of this approach in a polar ice sheet setting. Unlike some regions where downscaling methods, including ANNs, have been applied [e.g., central Europe; Huth (1999)], topography (especially in the interior) and seasonality are much less of an issue on the West Antarctic ice sheet, although they are still a factor in predictive skill. In particular, the inability of forecast models to capture the characteristics of the surface temperature inversion over the ice sheet may be a significant contributor to the accuracy of our temperature predictions. Thus the reduced meteorological complexity of a large ice sheet (away from the coast and mountains) is offset by the reduced skill of forecast models in this region further highlighting the need for better instrumental records.
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 European Centre for Medium-Range Weather Forecasts (ECMWF) surface data.
a. 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; that is, accuracy decreases with decreasing temperature (M. Lazzara 2001, personal communication). The data used here are from the 3-hourly quality-controlled datasets available at the University of Wisconsin—Madison Web site (http://amrc.ssec.wisc.edu/aws). A 6-hourly subset of these data (for 0000, 0600, 1200, and 1800 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 (Fig. 1). Ferrell was selected here because it has a longer and more continuous record (Fig. 4) than most AWSs, which are generally more remote and harder to service. The 18-yr (1981–98) average availability for temperature and pressure data, at 3-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 (1979–93), presumably from long-term equipment problems, occur in the late austral winter/spring during 1983–85 and 1991–92 (Fig. 4).
b. ECMWF data
The ECMWF 15-yr reanalysis data product (ERA-15) provided GCM-scale meteorological data for the period 1979–93 (ECMWF 2001a). 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 0000, 0600, 1200 and 1800 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.
1) 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 points plus points from the corners of a square area centered approximately on the station (Fig. 5). Ferrell is located fortuitously close to an ECMWF grid point. Testing showed that factors other than the gridpoint configuration have a substantially larger influence on performance.
2) 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., Fig. 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 (e.g., an elevation error for Vostok station has broad effects on geopotential heights). Evaluations of several operational products (e.g., Bromwich et al. 1995, 2000; Cullather et al. 1998) and discussions with experienced polar meteorologists (D. Bromwich and J. Turner 2000, personal communication) suggest that the ECMWF analyses are the best datasets 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-yr reanalysis is readily available (ECMWF 2001b).
At the simplest level, ANNs are a computer-based problem solving tool inspired by the original, biological neural network—the brain. Because of their ability to generate nonlinear 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 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.
a. ANN architectures
Three ANN types were used, all variants of the basic multilayer feed-forward ANN (Fig. 2). All share the same general form of processing node (Fig. 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. The FF ANNs consist of a large number of highly interconnected, simple processing nodes (a.k.a. neurons) organized into at least three layers (Fig. 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 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 nonlinear, 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 are 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).
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.
b. 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.
1) ANN training
The 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 back-propagation 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.
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 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 five best ANNs from the overall best performing configuration.
3) 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 gridpoint 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.
Our current ANN training methodology uses one calendar year of AWS observations to train each instance of the ANNs. This has yielded excellent performance within the training year, particularly for pressure, but reduced predictive skill in the other years. This approach is 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 for the most recently installed AWSs. Few of the existing AWSs have the record length available at Ferrell (1981–present) so it is useful to test the methodology under worst-case conditions. We also initially believed that 1 yr of training data would be sufficient to obtain acceptable predictive skill and this has been true within reasonable error limits. But it is also reasonable to think that performance could improve by taking advantage of more training data when it is available since the ANN would be “seeing” a wider variety of meteorological situations. This will be explored in future work.
Figure 6 graphically summarizes surface temperature prediction results for a training year (1987) and an arbitrary nontraining 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 Fig. 8. Although most nontraining years have lower performance than the training year results, there are some nontraining years that do nearly as well predicting when compared with the training year ANNs. Also shown in Fig. 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 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 nontraining 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 a distinct difference between higher summer and lower winter predictive skill. Spring and fall are nearly identical to the average rms error. The reduced skill in winter may be related to higher variability compared to summer (Table 2 and Fig. 4). This explanation is challenged by the higher variability but lower rms error of the spring season. Thus while there may be some relationship between season and predictive skill for temperature, it is not simple to define.
b. New records
After completing the ANN training process, the best ANNs were used to synthesize 15-yr 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 Fig. 9.
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. Unfortunately, very little on this subject in Antarctica has been published. 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.
a. 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 AWSs in West Antarctica. In the satellite-based methodology, 3-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 6-hourly AWS observations (for ECMWF time steps at 0000, 0600, 1200 and 1800 UTC) thus yielding up to four predictions each day (for those days with no missing 6-hourly AWS observations). Thus both our observed and calculated daily means are based on up to four 6-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 in the higher temperature region of this plot, which is likely related to the seasonal skill differences described earlier. A closer examination using a probability density function plot (not shown) 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 of SS's results (their Fig. 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's 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 in the training year is improved over the SS methodology (mean of 0.01° versus 0.14°C and standard deviation of 3° versus 5.5°C). For the mean values in nontraining years, the mean error in our results is larger (0.56° versus 0.14°C) but the mean standard deviation (σn−1) is nearly identical (5.48° versus 5.52°C). Examination of the ranges suggests that while our mean errors may have a larger range of values, the distribution around each mean is comparable to or tighter than what the SS methodology produces. 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 through other avenues such as refining the method used for selection of training records or training with seasonal data.
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, their Fig. 9). This type of error might be reduced by including 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; Fig. 11b shows all years (1981–93). 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.
b. 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 (Fig. 5). A comparison of 2-m temperatures from this grid point to the 2-m temperatures observed at the AWS (Fig. 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 (Fig. 10), the rms error is approximately twice as large (10.3° 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 (Fig. 13) and are comparable to our ANN-based predictions (Fig. 14). Our rms error (2.9 versus 4.7 mbar) and mean error (0.41 versus 4.13 mbar) are better, suggesting a benefit, albeit possibly slight, to our methodology for this variable.
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.
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. Finally, we will be able to extend our methodology into the presatellite era once the ECMWF 40-yr reanalysis datasets 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 AWSs with short observational records. This should also be the worst case for those AWSs 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.
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 RBA. We are also grateful to the Antarctic Meteorological Research Center, University of Wisconsin, for their archive of Antarctic AWS data.
Corresponding author address: David B. Reusch, Dept. of Geosciences, The Pennsylvania State University, 517 Deike Bldg., University Park, PA 16802. Email: email@example.com