1. Introduction

Antarctic climate variability and climate change in Antarctica have been based on near-surface temperatures measured at staffed weather stations (e.g., Jones 1995; Jacka and Budd 1998; Turner et al. 2005, 2016, 2020; Clem et al. 2020). While the annual mean temperature on the Antarctic Peninsula and some of the coastal regions of the continent is reported to have increased since the second half of the twentieth century (e.g., Turner et al. 2020), the figures are based on measurements taken mainly at staffed stations. Because of the remoteness and harsh environment of the region, only two staffed stations (Vostok and Amundsen–Scott) have been established in the interior Antarctic Plateau. In particular, the western sector of the East Antarctic Plateau, that is, interior Dronning Maud Land (DML), has no staffed stations. As the network of stations is insufficient to cover the entire continent, DML has been too data sparse to permit studies of its temperature trends.

Some of the gaps in the network of stations have been filled by the installation of automatic weather stations (AWSs). In DML specifically, the Japanese Antarctic Research Expedition (JARE) installed its first AWS along the route to Dome Fuji, 1000 km inland from the coastal area, in the early 1990s, and has been updating the station and replacing its instruments up to present. The first AWS system, which recorded data in complementary metal oxide semiconductor (CMOS) memory and operated at very low temperatures with low power consumption (hereafter referred to as “CMOS-AWS” systems), was developed by the Antarctic Climate Research (ACR) program between 1987 and 1992 (Kikuchi and Endoh 1993). In 1993, CMOS-AWS systems were installed at three sites, namely, Mizuho, Relay Station, and Dome Fuji (see Fig. 1), to measure air temperature as part of the Deep Ice Coring Project at Dome Fuji (Enomoto et al. 1995; Kameda et al. 1997; Takahashi et al. 1998). The three stations continued their observations throughout the duration of the project and were completed in 2010. In addition to the CMOS-AWS systems, four AWS units upgraded with improvements by the Antarctic Meteorological Research Center (AMRC) at the University of Wisconsin–Madison (hereafter UW-AWSs), were also deployed in the DML interior (Mizuho, Relay Station, JASE2007, and Dome Fuji). UW-AWS observations began in collaboration with the U.S. Antarctic Program at Relay Station and Dome Fuji in 1997, at Mizuho in 2001, and at JASE2007 in 2007 (Lazzara et al. 2012). Much progress has been made in recent decades in developing meteorological instruments capable of providing more accurate temperature data. The latest AWS systems were introduced by the JARE at H128 in 2017, at NDF and Relay Station in 2018, and at MD78 in 2019. Advances in the AWS network have helped fill a major gap in temperature data for the DML interior. Combining these AWS observations creates a dataset that allows for the study of interannual variability and surface air temperature trends in the region.

View Full Size

A map of east Dronning Maud Land in East Antarctica with locations of AWS observations (squares) along the route to Dome Fuji. The blue line represents the DF route. Inset: A map of Antarctica showing the study area. Red squares represent the locations of the three AWS sites that have been in operation for 30 years. The other AWS and meteorological stations are shown as black squares.

Citation: Journal of Atmospheric and Oceanic Technology 41, 2; 10.1175/JTECH-D-23-0092.1

• Download Figure
• Download figure as PowerPoint slide

Several concerns need to be addressed, however, before the AWS observations can be used for the wider study of climate change. One concern is that temperature measurements in the Antarctic Plateau may have been biased warm in the austral summer (e.g., Genthon et al. 2011; Morino et al. 2021). On the plateau, intense solar radiation and low wind conditions result in insufficient ventilation of the temperature sensor in its radiation shield, leading to a warm temperature bias. A second concern is the change in the sensor height due to snow accumulation. Due to the strong surface temperature inversions in the remote interior, gradual decreases in the sensor height lead to a cold bias. A third concern is the possible occurrence of systematic errors in the records stemming from changes in the instrumentation at the stations. All of these factors make it difficult to detect the true signals of climate change. In this study we quantify the errors associated with these factors to create a corrected dataset that is available for the study of climate variability and change in the DML interior. In Antarctica, the Scientific Committee on Antarctic Research (SCAR) Reference Antarctic Data for Environmental Research (READER) project has created and distributed the most reliable and consistent climate dataset (Turner et al. 2004). The READER dataset includes monthly and annual mean near-surface temperatures from the UW-AWS records in DML. Errors associated with the above three factors have not been completely eliminated, however, in spite of the quality-controlled UW-AWS data provided to the READER project. The primary goal of our study is to recompute the monthly mean climate data from all of the available AWS observations in the DML interior. Monthly means are computed following the data-processing steps used for READER. The READER dataset only includes AWS data collected from stations that have either been in operation for the last 5 years or that were operated for a period of at least 10 years at some point in the past. Here, therefore, we present a dataset from three AWSs that satisfy these conditions (Mizuho, Relay Station, and Dome Fuji).

The next section presents the instrumentation and AWS observations in the DML interior. The methods for correcting the radiation bias and producing a single time series are explained in sections 3 and 4, respectively. Section 5 compares the new products with the reanalysis data. Finally, the general conclusions are summarized in section 6.

2. Data

3. Observed radiative errors and corrections

Figure 2 shows the differences between the temperature measured by the FV and NV sensors during the austral summer at Dome Fuji. Differences of less than 2°C were judged to be noise in this study, and our previous study demonstrated that the differences between the two sensors varied randomly within ±2°C during the polar night (Morino et al. 2021). Differences of more than 2°C, which are considered radiation bias, were often observed in both the CMOS-AWS and UW-AWS measurements. The magnitude of the bias, however, differed between the AWS observations, as the CMOS-AWS errors were approximately twofold greater than those of the UW-AWS. The hourly temperature errors reported by both the CMOS-AWS and UW-AWS systems clearly reflected a diurnal cycle of solar radiation. The errors began to increase from local morning and peaked at local noon, at some points reaching temperature differences above 10°C. The radiation bias varied considerably with the shape and material of the NV shield. We previously found that the radiation bias measured with the UW-AWS shield was much lower than that with the standard NV shield (Morino et al. 2021). Based on the Morino et al. (2021) study, we surmise that the NV shield used for the UW-AWS absorb less radiant energy and/or has a higher ventilation efficiency than the NV shield used for the CMOS-AWS.

View Full Size

Histograms of the frequency distribution of radiation error (difference between the FV and NV sensors) for (a) CMOS-AWS and (b) UW-AWS.

Citation: Journal of Atmospheric and Oceanic Technology 41, 2; 10.1175/JTECH-D-23-0092.1

• Download Figure
• Download figure as PowerPoint slide

As noted in Morino et al. (2021), the radiation bias in the interior DML increases linearly with shortwave radiation and decreases nonlinearly with wind speed. For this reason, the radiation bias can be estimated using the regression model proposed by Nakamura and Mahrt (2005). This model expresses the radiation bias as a function of a dimensionless variable: the ratio of the heating by solar radiation to the cooling by natural ventilation. Here we calculate a dimensionless variable X, as follows:

$X=\frac{{\text{SW}}_{\text{ref}}}{\rho C_{p}TW},$

where SW_ref is reflected shortwave radiation from ERA5, ρ is air density (kg m⁻³), C_p is specific heat capacity of air (J kg⁻¹ K⁻¹), and T (K) and W (m s⁻¹) are the measured air temperature and wind speed, respectively. Figure 3 shows comparisons of the radiation bias in the CMOS-AWS and UW-AWS measurements with a dimensionless variable. We developed separate correlation models for the CMOS-AWS sensor and the UW-AWS sensor, as each has a unique regression according to the sensor/shield combination. Note that the regression curve for the UW-AWS sensor obtained from the observations in the 2000s significantly differed from that obtained from the recent observations (orange plots in Fig. 3b), with an increase in radiation bias effects seen in the latter. The higher bias may have been due to an aging effect, as the surface of the radiation shield degrades over time through constant exposure to sunlight and weather (Lopardo et al. 2014). We used three different regression models for the UW-AWS data in this study, depending on the observation period. For the period with no FV sensor measurement (2007–17), we used the intermediate values to reduce the radiation bias between the estimates from the model before 2006 and after 2018. Figure 4 shows the robustness of this approach. The warm bias seen in the original data during the austral summer was successfully reduced through the use of the regression model. The root-mean-square error (RMSE) was reduced more than 60% for both AWS sensors. Although the RMSE of the CMOS-AWS (1.4°C) was greater than that of the UW-AWS (0.8°C), the RMSE of the former was comparable to the accuracy of the measurement.

View Full Size

Radiative errors ΔT (NV − FV) as a function of the dimensionless variable (the ratio of heating by solar radiation to the cooling by natural ventilation) for the (a) CMOS-AWS sensor and (b) UW-AWS sensor at Dome Fuji in austral summer. For a better display, the dimensionless variable is multiplied by a factor of 10³. The solid blue line represents the regression curve before the year 2007, when the JMA-AWS observations were completed. The orange plots in (b) are for the UW-AWS sensor at Relay Station after the year of 2018. The solid red line is the regression with plots after 2018.

Citation: Journal of Atmospheric and Oceanic Technology 41, 2; 10.1175/JTECH-D-23-0092.1

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Temperature measured from the FV sensor vs the original (light blue plots) and corrected temperatures (orange plots) from the NV sensors installed at the (a) CMOS-AWS and (b) UW-AWS at Dome Fuji in austral summer. (c) As in (b), but for the UW-AWS at Relay Station between 2018 and 2022. The broken line represents 1:1. The root-mean-square error (RMSE) between the FV temperature and corrected NV temperature is also shown.

Citation: Journal of Atmospheric and Oceanic Technology 41, 2; 10.1175/JTECH-D-23-0092.1

• Download Figure
• Download figure as PowerPoint slide

4. Merging of the AWS records

Figure 5 shows comparisons of the warm-bias-corrected temperatures (T_corr) reported by the UW-AWS and CMOS-AWS sensors at the three stations. Some of the notable differences between the two records arose not from sensor error, but from inconsistency in the datalogger clocks between the two AWS sensors. As the CMOS-AWS logger time constantly lagged behind, the clock in the logger preceded the actual time by a wide margin of several hours to almost a day when we visited a year later. We corrected for the time by assuming a linear delay, which caused some uncertainty in the observation times recorded in the CMOS-AWS data. An error of a few hours in the corrected time resulted in large temperature differences over the Antarctic Plateau, a region with a distinctive diurnal cycle. The temperature differences at Mizuho showed a strange distribution (Fig. 5c). From the time series, we see that the pattern of the diurnal cycle and day-to-day variability reported by the UW-AWS and CMOS-AWS sensors was moderately consistent (not shown). The variability in the UW-AWS data, however, far exceeded that in the CMOS-AWS data and presented as an oddly shaped sawtooth wave. In addition, the maximum peaks of T_corr occasionally rose to near melting point (0°C). These patterns indicate that the temperature data from the UW-AWS sensor were measured incorrectly for more than decade from the time the instruments were installed in 2001. Later, however, in January 2014, the sawtooth-like variability of the UW-AWS data suddenly disappeared. From that point forwarded we see smooth changes in the diurnal cycle and daily variability reported by the UW-AWS, and the summer temperatures no longer reached levels approaching 0°C. Therefore, we decided to use the UW-AWS reports from Mizuho only after 2014. As the differences reported by the other two stations showed no clear temperature dependence and were relatively uniform in distribution, the instrumental error due to differences in height between the two sensors appeared to be negligible. Low-level temperature inversions are a well-known feature in the interior of Antarctica, and the strength of the inversions is greater in austral winter than in austral summer (e.g., Pietroni et al. 2014). These patterns suggest that a temperature sensor close to the surface records a colder temperature and that the difference between the sensors increases from the austral summer to winter. A relatively uniform distribution throughout the year indicates that the bias due to the different heights of the AWS is negligible. The reported temperature from the UW-AWS sensor at Relay Station was consistently about 2°C lower than that from the CMOS-AWS sensor. Contrary to expectation, the UW-AWS sensor height was much larger than that of the CMOS-AWS sensor (2.0–3.0 m above the surface versus 1.2–1.5 m). This difference in sensor height above the surface further supports our conclusion that the instrumental error arising from the height difference was insignificant. Table 1 shows the manual surface air temperature observations with the UW-AWS and CMOS-AWS measurements at Relay Station in early spring, a time of year when solar radiation is very weak. The CMOS-AWS sensor data agree well with the manual observations within the measurement error. While the UW-AWS sensor data also reflected the daily variations, the recorded temperatures were about 2°C lower than those recorded manually. While this pattern suggests that the UW-AWS sensor at Relay Station had a significant cold bias of about 2°C, the bias was only observed in the measurements taken with the old AWS-2B system. After the replacement of the AWS-2B with a new CR1000-based datalogger system in January 2010, the UW-AWS sensor data agreed well with the latest AWS sensor data and had a smaller RMSE (0.44) (see Fig. 6). The AMRDC reported that the AWS-2B temperature observations at Byrd in West Antarctica had a calibration error of 1.5°C (Bromwich et al. 2012). A postinspection of Relay Station AWS-2B data by AMRDC, however, uncovered no signs of software or hardware troubles that could have resulted in calibration error. The source of the offset therefore remains unknown. The temperature drift in the low temperature range reported in Bromwich et al. (2012) was also observed at Dome Fuji, but not at Relay Station. At Dome Fuji, the temperature drift was observed at temperature below −50°C. The temperature error at Dome Fuji increased at lower temperature, reaching a 2°C margin at −80°C. Once these systematic biases in the UW-AWS sensor records were corrected, the bias between the UW-AWS and other AWSs sensors was reduced and the temperature differences were almost identical to the measurement error (the RMSE was 1.44°C for Dome Fuji and 0.88°C for Relay Station). These small values suggest that local environmental effects, such as unusual atmospheric and microclimatic effects, had no significant impact on our AWS observations. In this study, therefore, we merged the revised UW-AWS records with the other AWS records to create a single time series of data at each AWS site.

View Full Size

Hourly temperature differences between the UW-AWS and CMOS-AWS sensors at (a) Dome Fuji, (b) Relay Station, and (c) Mizuho. The RMSEs for the sensors are also shown.

Citation: Journal of Atmospheric and Oceanic Technology 41, 2; 10.1175/JTECH-D-23-0092.1

• Download Figure
• Download figure as PowerPoint slide

Table 1.

Surface air temperature observed manually at Relay Station from 8 to 12 Sep 2000. The bracketed values represent the differences compared to manual observations.

View Table

View Full Size

As in Fig. 5b, but for (a) the differences between the AWS-CR1000 system and the latest AWS reports over the period 2018–22 and (b) the monthly mean temperature differences between the corrected AWS observations (AWS_corr) and READER data. Open circles represent the data from 1995 to 2010. Gray circles are the data between 2011 and 2022.

Citation: Journal of Atmospheric and Oceanic Technology 41, 2; 10.1175/JTECH-D-23-0092.1

• Download Figure
• Download figure as PowerPoint slide

5. Climate data processing

Quality-controlled hourly data allow the production of a more accurate series of monthly means. As with the READER data, the monthly means were computed using 6-hourly data rather than hourly AWS observations. To adjust for the time error of the AWS reports, the 6-hourly data were averaged to include observations from ±1 h (e.g., data at 0600 UTC were averaged from 0500 to 0700 UTC). If a value at a nominal time was missing, a 6-hourly value was estimated by averaging two adjacent values, or each adjacent value was substituted. Following the threshold used in the READER data, we only computed a monthly mean if 90% or more of the 6-hourly data were available and there were no gaps of more than two consecutive days. Figure 7 shows the time series of corrected temperatures from the AWS observations at Mizuho, Relay Station, and Dome Fuji over the past 30 years. The remoteness and harsh climatic conditions in the interior of Antarctica prevented frequent maintenance visits. As we were unable to correct for instrumental failures and power failures, the observations could not be captured as a continuous record. The percentage of missing data was 25% at Relay Station, and 41% at Dome Fuji due to the even harsher environment. By adjusting the cutoff percentage for the computation of the monthly mean from 90% to 70%, however, the percentage of missing data decreases to 21% at Relay Station and 28% at Dome Fuji. The rate of missing data at Mizuho, where no UW-AWS temperature data were available until January 2014, was about 49%, or about double the rate at the other two stations. We therefore examined the impact on the monthly mean temperature when the averages were computed from data with a low cutoff of 70%. The monthly mean temperature was recalculated by randomly selecting 70% of the data from the whole month and comparing it with the original data using a 90% cutoff. The RMSE calculated for each month throughout the entire observation period is shown in Figs. 8d–f. Because of the shorter period with available data, the RMSE is higher for Mizuho than for the other sites in every month. In Relay Station and Dome Fuji, the RMSE increased relative to the monthly mean temperature (shown in Figs. 8a–c) as the temperature decreased. We find, however, that the RMSE values were almost less than a third of the standard deviations of the monthly temperatures (Figs. 8d–f). The annual mean RMSE is approximately 0.6°C. This level of error is also sufficiently small when compared to the variation in the annual mean at the three sites. From these results, we can say that the 20% increase in the missing data rate seems to have no significant effect.

View Full Size

Monthly mean temperature time series from the corrected AWS observations (red line) and monthly mean 2-m temperature from ERA5 (dashed gray) at (a) Dome Fuji, (b) Relay Station, and (c) Mizuho. The orange line is the corrected AWS observations, but with the cutoff percentage for the computation of the monthly mean reduced to 70%.

Citation: Journal of Atmospheric and Oceanic Technology 41, 2; 10.1175/JTECH-D-23-0092.1

• Download Figure
• Download figure as PowerPoint slide

View Full Size

(top) Long-term mean (1993–2022) of observed monthly average temperature at (a) Dome Fuji, (b) Relay Station, and (c) Mizuho. The error bars indicate the standard deviations of each month throughout the observation period. (bottom) Histograms of the RMSE (gray bars) of monthly mean temperatures from the 70% cutoff percentage for the computation are shown at (d) Dome Fuji, (e) Relay Station, and (f) Mizuho. The standard deviations in (a)–(c) are also shown (white bars).

Citation: Journal of Atmospheric and Oceanic Technology 41, 2; 10.1175/JTECH-D-23-0092.1

• Download Figure
• Download figure as PowerPoint slide

6. Comparison with other data

The READER project reports the monthly mean temperature from the UW-AWS at Relay Station. Figure 6b compares the monthly means reported by READER with the corrected values in this study. READER is based on 6-hourly observations that are uncorrected for the offset of the AWS-2B and radiation-induced warm bias. This led to notable biases, especially before 2010, when the READER data consistently showed a cold temperature bias greater than 2°C, and in the summer after 2010, when the READER data showed a warm temperature bias of approximately 2°C.

A comparison of corrected AWS data with ERA5 monthly mean 2-m temperature data is shown in Fig. 9. The ERA5 data agree well with the warm-bias-corrected temperature in summer but overestimate the bias in the other seasons. The temperature differences in the winter reach up to 5°C at Relay Station and Dome Fuji. These differences are consistent with the comparisons with other inland observations (Gossart et al. 2019). In addition, we find no systematic differences at Relay Station attributable to the influence of the AWS-2B offset correction. It is well known that temperature observations from AWSs are used to produce reanalyses through many data assimilation systems. Here, however, we speculate that the AWS observations in the DML interior have no significant impact on ERA5. As expected, the distribution of the temperature differences for the cutoff percentage of 70% corresponds well to that calculated with a data availability rate of 90%.

View Full Size

Differences between corrected AWS reports and 2-m temperature from ERA5 at (a) Dome Fuji, (b) Relay Station, and (c) Mizuho. Red plots are monthly values calculated with a data availability rate > 90%. Open orange circles are the same as the red plots, but with a cutoff percentage of 70% applied. The RMSEs for these plots are also shown. The bracketed values represent the RMSEs for the 70% cutoff percentage.

Citation: Journal of Atmospheric and Oceanic Technology 41, 2; 10.1175/JTECH-D-23-0092.1

• Download Figure
• Download figure as PowerPoint slide

7. Conclusions

We present a new dataset of monthly mean near-surface air temperatures derived from historical AWS observations in the DML inland of East Antarctica over the past 30 years. While no staffed weather stations are in place in the DML interior, several AWSs were installed along the route to Dome Fuji in the 1990s, and observations have continued at three stations (Mizuho, Relay Station, and Dome Fuji) to the present day. AWS observations in the plateau region are significantly affected by systematic errors caused by solar radiation over the observation period. There were also measurement errors and drift in the AWS system in the early days. In this study we quantify how the errors and drift affect the AWS observations, and apply corrections to remove their effects. A comparison with the data from READER at Relay Station shows significant differences between the READER data and our own. The differences suggest that our data are more accurate than the previously available data, as the latter contain systematic errors. The harsh environments in which the AWSs are usually installed have made it possible to collect continuous records over longer periods of decades. In this study, however, we integrated multiple AWS records to create a near-continuous record from 1993 to 2022. The percentage of missing data over the three decades was 21% for Relay Station, 28% for Dome Fuji, and 49% for Mizuho. Missing observations can be estimated using temperatures from various reanalyses, which do not use any AWS observations in their data assimilation system (e.g., Bromwich et al. 2012). ERA5 is one of the candidates, as the monthly 2-m temperature data from ERA5 do not respond to the systematic biases that affect the original AWS observations and appear to be independent of the respective datasets analyzed. The reconstructed continuous records allow us to study temperature variability and change over the western sector of the East Antarctic Plateau, where climate change has so far been unexplored.

Data availability statement.

The CMOS-AWS data are distributed by CD-ROM from the Kitami Institute of Technology. The JMA-AWS data are available at the Antarctic Meteorological Data online archive (https://www.data.jma.go.jp/antarctic/datareport/index.html). The latest AWS data conducted by JARE are available through the Arctic and Antarctic Data archive System (AADS) operated by the National Institute of Polar Research. The other AWS data can be accessed through the Antarctic Meteorological Research and Data Center (AMRDC) Data Repository at https://doi.org/10.48567/1hn2-nw60 or https://amrdcdata.ssec.wisc.edu/group/automatic-weather-station-project. The READER data are available from the https://legacy.bas.ac.uk/met/READER/aws/awspt.html.