China not only has a vast land area covering tropical, temperate, and sub-frigid zones, but also has a complicated topography with the average elevation of the Tibetan Plateau more than 4000 m above the sea level. Among the 114 radiosonde stations over China, 16 of them have a surface elevation greater than 2000 m. Of these 16 stations, most are located in the Tibetan Plateau (Fig. 1). It is of interest to understand how the MSU and radiosonde data are correlated in such a complicated topography. Note that for the high mountain and plateau areas, it is ideal to use MSU weighting functions specifically calculated for these areas to convert radiosonde data to MSU equivalents. However, such weighting functions are unavailable and thus we used the standard MSU weighting functions that are based on the standard atmosphere for conversions from radiosonde to satellite equivalents for all stations over China. Alternatively, using radiative transfer models to calculate MSU equivalent temperatures is a viable approach for both plain and plateau areas and worth carrying out in future studies.
Figures 4 and 5 show correlations and root mean square deviation (RMSD), respectively, between the MSU datasets from different groups (STAR, RSS, and UAH) and ADJ for different layer temperatures. Statistically significant positive correlations greater than 0.9 and consistent RMSD are found for TLS over the entire China for all the different MSU datasets. This suggests that the weighting function for TLS, which peaks near 70 hPa, is high enough so that the high-altitude Tibetan Plateau does not have an effect on the radiosonde conversion. For TUT, the relatively lower correlation and larger RMSD are found over southern China and Tibetan Plateau for all the satellite datasets. This is related to radiosonde data issues near the tropopause in these areas as discussed in the previous section. Given that its weighting function peaks (near 250 hPa) higher than the Tibetan Plateau, the topography has a negligible effect on the radiosonde equivalent for TUT.
Figure 4. Correlation of monthly temperature anomalies between MSU (STAR, RSS, and UAH) and ADJ over China during 1979–2015. So-lid circles indicate the st atistical significance of correlation (> 0.35). Panels from top to bottom are for TLS, TUT, TMT, and TLT; panels from left to right are for STAR, RSS, and UAH.
Figure 5. RMSD of monthly temperature anomalies between MSU (STAR, RSS, and UAH) and ADJ over China during 1979–2015. Panels from top to bottom are for TLS, TUT, TMT, and TLT; panels from left to right are for STAR, RSS, and UAH.
For TMT, the large correlation greater than 0.9 and consistent RMSD are found over most regions in China for all the three MSU datasets except over the Tibetan Plateau. This occurred because the MSU TMT weighting function peaks at about 500 hPa, close to the surface elevation of the Tibetan Plateau. This weighting function is not expected to work well for converting the radiosonde data to TMT equivalent over the Tibetan Plateau. Larger differences, as represented by RMSD over most Tibetan Plateau regions (Fig. 5), between the MSU and radiosonde equivalent based on such a weighting function inevitably occur. For TLT, high correlations above 0.9 are found for both RSS and UAH, but RSS RMSD is apparently larger than those of UAH for most regions. Relatively smaller correlations and larger RMSD are found over the Tibetan Plateau for both UAH and RSS. Similar to TMT, this is expected as the MSU TLT weighting function peaks below the surface of the Tibetan Plateau where it is inappropriate for conversions from radiosonde level temperatures to satellite layer temperatures. Hence, it is better to remove these data for more accurate trend estimates over China. It is seen that beyond 2000 m the correlation decreases and RMSD deceases more rapidly with the surface elevation; as a result, the 2000-m elevation is a reasonable criterion to separate stations for surface elevation effects. In the following discussion, regional mean anomalies and trends over China are generally calculated over areas where surface elevations are less than 2000 m (hereinafter being referred to as Cn-notb in Tables 1, 3), although tests have been conducted to include stations with elevations above 2000 m (Cn, Cn-sta, and Cn-nosta in Table 3).
Item Dataset TLT TMT TUT TLS Correlation MSU, ADJ 0.96 0.92 0.73 0.96 MSU, HAD 0.95 0.92 0.86 0.92 MSU, RAW 0.96 0.86 0.56 0.93 MSU, ReADJ 0.96 0.96 0.91 0.97 Standard Deviation MSU 0.68 0.50 0.39 0.95 ADJ 0.59 0.39 0.36 0.86 HAD 0.61 0.43 0.38 0.73 RAW 0.61 0.40 0.41 0.98 ReADJ 0.60 0.40 0.34 0.83 RMSD MSU–ADJ 0.19 0.16 0.25 0.25 MSU–HAD 0.22 0.17 0.19 0.32 MSU–RAW 0.19 0.21 0.35 0.38 MSU–ReADJ 0.19 0.12 0.14 0.20
Table 1. Comparison statistics for monthly layer temperature anomalies between MSU and radiosonde datasets averaged over China (Cn-notb) during 1979–2015 (STAR dataset is used for MSU TMT, TUT, and TLS statistics while the RSS dataset is used for MSU TLT statistics)
Item TLT TMT TUT TLS Globe 0.117 ± 0.051 0.111 ± 0.051 0.022 ± 0.060 –0.278 ± 0.242 Cn 0.186 ± 0.075 0.188 ± 0.058 0.055 ± 0.050 –0.346 ± 0.168 Cn-sta 0.174 ± 0.072 0.176 ± 0.056 0.056 ± 0.046 –0.336 ± 0.165 Cn-nosta 0.199 ± 0.080 0.199 ± 0.062 0.056 ± 0.054 –0.355 ± 0.168 Cn-notb 0.149 ± 0.064 0.134 ± 0.044 0.040 ± 0.039 –0.278 ± 0.135 ReADJ 0.203 ± 0.066 0.128 ± 0.044 0.034 ± 0.039 –0.329 ± 0.135
Table 3. Linear trends (K decade–1) of monthly layer mean temperature anomalies during 1979–2015 for ReADJ and MSU averages over the globe and China from STAR, RSS, and UAH datasets. Cn denotes satellite averages over the entire China, Cn-notb for areas without the Tibetan Plateau, and Cn-sta and Cn-nosta for areas with and without stations. Uncertainties represent the 95% confidence level with autocorrelation included
Table 1 summarizes the comparison statistics associated with findings in the previous subsection in terms of the correlation and RMSD between MSU and the radiosonde datasets before (RAW) and after homogenizations (ADJ and HAD) during 1979–2015. Statistics are for stations in China without the station elevation above 2000 m (Cn-notb). The summary emphasizes common features in the differences and agreement between the satellite and radiosonde data before and after the homogenization. The STAR data are used to represent satellite MSU data for TMT, TUT, and TLS, while the RSS data are used for TLT. Statistics for the same layer temperatures but from UAH and RSS datasets show similar results, although their correlations with the radiosonde data are slightly higher for certain layer temperatures. This will not affect discussion on differences between the radiosonde and MSU data.
The results revealed the excellent agreement between MSU and homogenized radiosonde observations with statistical significant positive correlations greater than 0.9 and similar RMSD for TLT, TMT, and TLS. Correlations for TUT are lower than those of the other layer temperatures. The radiosonde homogenization and readjustment obviously improve the agreement of ADJ and ReADJ (the re-adjusted ADJ time series, details in Section 4.2) with MSU data. Both ADJ and HAD show a better agreement with the MSU dataset than RAW by the remarkable higher correlation coefficient and smaller RMSD for TMT, TUT, and TLS. For TMT, MSU correlations go from 0.86 with the RAW up to 0.92 with the ADJ (HAD), 0.96 with ReADJ. Similar improvements are also found for TUT. As discussed in Fig. 6, this resulted from reduction of the discontinuity in raw radiosonde temperatures. On the other hand, correlations between MSU and radiosonde data remain the same for TLT before (RAW) and after (ADJ and HAD) homogenizations. The reason for this is that the impact of homogenizations for TLT is generally much smaller than the other three layers (Chen and Yang, 2014). The reason why HAD agrees better than ADJ with MSU is the impact of different references used in homogenizations (Guo and Ding, 2011). Since radiosonde observations are an input source for reanalysis, the independence between the RAW and reference causes the inhomogeneity in nighttime temperature time series hard to be detected or adjusted thoroughly with the reference from ERA-Interim at individual stations. While the reference of HAD is based on the larger scale network and thus effectively minimizing the large systematic biases in raw data (Thorne et al., 2005b). This approach complements the deficiency in ADJ reference and, as a result, improves the consistency with MSU.
Figure 6 presents annual temperature anomalies averaged over China during 1979–2015 from radiosonde and MSU datasets (Figs. 6a1–d1) and their differences from the RAW (Figs. 6a2–d2). Note that the use of RAW as a common reference for differencing here is to help assess the impact of homogenizations from HAD and ADJ. Annual anomalies show a good agreement between the MSU datasets and radiosondes in terms of similar variability. The time series of differences shows more diversity in TUT and TLS than that in TLT and TMT. Comparing with MSU anomalies, RAW anomalies show an apparent sharp decrease nearly 0.5 K in 2001, causing the entire time series of differences, which are divided into negative and positive phases across 2001. This large shift is closely related to the radiation correction and radiosonde system update in China during 1999–2001 as well as the replacement of GZZ2 with GTS1 (L-band) during 2002–10 over China reported in radiosonde metadata records (Chen and Yang, 2014; Guo et al., 2016). Both HAD and ADJ have detected the shift around 2000 and adjusted the RAW to a certain extent. Meanwhile, MSU anomalies are smoother around radiosonde jumping points, indicating that MSU data could be used as potential references to validate original radiosonde temperatures. Comparing with ADJ, HAD adjusts the RAW more substantially (larger difference with RAW) and has better agreements with MSU datasets. This suggests that ADJ may have residual discontinuities and HAD is more reliable especially during the 2000s.
The discontinuity problem shown in regional mean anomalies can be better illustrated by anomalies at individual stations. This is because the station metadata, with exact information on the station location and instrument changes, are more accurate for quantifying the artificial influence caused by observing system changes. As an example, monthly temperature anomalies from ADJ, HAD, and the three MSU datasets at station 52681 (which has the most complete record) are shown in Fig. 7. Although such a time series is noisier than that of the annual and regional averages, stepwise changes and timing in the anomalies match very well with the radiation correction and radiosonde system updating during 1999–2001 and instrument change from GZZ2 to GTS1 in January 2006 at this station. Both ADJ and HAD have detected changing points and adjusted RAW with different magnitudes. Similar to Fig. 6, HAD adjusts RAW with the larger magnitude than ADJ since the former has larger differences from RAW in 2001. Generally, HAD has a better agreement with MSU datasets than ADJ in TLT, TMT, and TUT (Table 1), due to different homogenization procedures. In contrast to the obvious discrepancy between HAD and ADJ, MSU datasets show remarkable high consistency in their differences from RAW with the similar phase and magnitude in variability. This suggests that MSU data from different groups either all represent the reality very well or have common issues in their merging, especially during the period for radiosonde system updating or instrument change in the 2000s.
Figure 7. Monthly temperature differences (K) with respect to RAW at station 52681 (Minqin) from ADJ, HAD, and the three MSU datasets (STAR, RSS, and UAH) during 1979–2015.
Five-year moving trends of the monthly differences at individual radiosonde stations can reveal more information on timing when differences between the MSU and radiosonde occur (Randel and Wu, 2006; Randall and Herman, 2008; Mears et al., 2012). Figure 8 shows 5-yr moving trends of differences at four different stations for TUT between the data pairs. These four stations are representatives of the northwestern, northeastern, northern, and eastern China (52681 Minqin, 50953 Harbin, 54511 Beijing, and 58362 Baoshan; Fig. 1). These 5-yr moving trends are composed of the maximum and minimum at different times, representing bias jumps between the data pairs at the specific time. Roughly speaking, maximum represents the time when ADJ, HAD or MSU data start to become warmer than RAW while the minimum is opposite. As a result, the maximum and minimum in Fig. 8 represent the most significant changing points between the data pairs. Table 2 summarizes the times when large changes occur at these stations between MSU and RAW data pairs. The maximum around 1999–2001 and mini-mum around 2004–06 occurred for all the four stations and they matched well with the two major changes in the radiosonde observations: system update over China during 1999–2001 and instrument mode replacement at station 52681 in January 2006, at station 50953 in January 2005, at station 54511 in January 2002, and at station 58362 in August 2003. Four major changes are identified in 1983, 1986, 1988, and 1991 for at least two of the four stations (Fig. 8). Metadata for these stations indicate that there are no major radiosonde changes during this period. This is why ADJ is close to RAW during this period for all the four stations. Consequently, these changes are most likely related to satellite changes. Indeed, the years for these changes respectively coincide well with the launch time of NOAA-8, NOAA-10, NOAA-11, and NOAA-12 satellites when MSU data onboard these satellites become available to merge with previous satellites for development of the MSU time series. These consistencies are also found for differences between the other MSU layer temperatures and RAW data. Such information suggests that radiosonde data at individual stations can help to identify bias jumps in the merged satellite time series, which will in turn help the satellite data developer to improve the merging accuracy of satellite CDRs.
Figure 8. Five-year moving trends for differences (K decade–1) of TUT with respect to RAW from ADJ, HAD, and the three MSU datasets (STAR, RSS, and UAH) during 1979–2015 at the four stations: (a) station 52681 (Minqin), (b) station 50953 (Harbin), (c) station 54511 (Beijing), and (d) station 58362 (Baoshan).
Year Possible cause Change occurred for the station number 1999–2001 Radiation correction and radiosonde system update over China 4 2002–2010 Change of radiosonde instrument models from GZZ2 to GTS1/L-band 4 1983 Start of NOAA-8 2 1986 Start of NOAA-10 3 1988 Start of NOAA-11 3 1991 Start of NOAA-12 3
Table 2. Times of major changes that occur at the four stations in China during 1979–2015 between MSU and RAW for TUT
Similar breakpoints are also found for differences between ADJ and RAW during 1999–2001 and 2002–10 for individual stations (Figs. 7, 8) and regional averages (Fig. 6), although magnitudes of the breakpoints vary. Given the fact that differences between MSU and RAW are much larger during the radiosonde transition period of 1999–2001 and differences between ADJ and RAW are much smaller during the same period (Fig. 6), it is likely that the adjustment in ADJ is not large enough to remove biases caused by the radiosonde system update. In other words, residual biases may still remain in ADJ after 2001. In order to reduce its impact on trends in the homogenized radiosonde data, we perform further re-adjustment on ADJ by using MSU averages over the three groups (abbreviated as MSUav) as a reference. We use a shift-point adjustment approach similar to those used in Christy et al. (2018) for re-adjustment. However, instead of conducting re-adjustment at each individual station, we only re-adjust the mean time series over China as a demonstration of the adjusting concept. In this approach, the maximum magnitude of difference shift between RAW and MSU datasets in Fig. 6 is firstly selected as shifting points, being 2001 for events of the 1999–2001 radiosonde system update and 2005 for events of 2002–10 radiosonde instrument changes. These shift points actually correspond to the maximum and mini-mum points for relevant events in the 5-yr moving trends as seen in Fig. 8, except that the 5-yr moving trends here should be for the averaged time series over China (not shown). After these shift points are determined, the regional mean ADJ is adjusted to match with MSUav at and after the shift points by the difference between two segments of 36-month in length on either side of the shift point from ADJ and MSUav. The re-adjusted ADJ time series, abbreviated as ReADJ, is also shown in Fig. 6. Comparison statistics (Table 1) show that ReADJ agrees much better with the MSU time series and its trends are discussed in Section 5.
Note that breakpoints during 1980–91, which are attributed to satellite launches in the MSU merging (Table 2 and Fig. 8), are not adjusted in this study. This is because these breakpoints caused the differences between MSU and radiosonde to jump up and down at different times (Figs. 6, 7). To the first order approximation, they did not appear to cause major trend differences between the two observational types because biases due to these up-and-down jumps cancelled out in the long-term trend estimates. This is different from radiosonde-related breakpoints, which cause stepwise changes in the time series of differences between satellite and radiosonde data. Such stepwise changes are the main reason causing the long-term trends from radiosonde data to be cooler than the satellite for TMT, TUT, and TLS and thus needing to be adjusted. The satellite adjustment will be conducted in future investigations when radiosonde is adjusted at individual stations.
By the inter-comparison between radiosonde equivalence and MSU layer temperatures, the significant difference in the 2000s over China has been found, which is related with residual discontinuities in ADJ even after homogenizations. For temperatures at neighboring stations are commonly used as the reference in homogenizations, we selected Hongkong (45004, 22.3°N, 114.2°E) as the reference of Qingyuan (59280, 23.7°N, 113.1°E) (Fig. 1) to check the discontinuity in the 2000s and MSU applicability as the reference. For station 45004 is not contained in Chinese radiosonde network, nationwide radiosonde system changes over China around 2000 have no impact on the temperature time series at station 45004. The layer mean temperatures from three MSU datasets and radiosonde equivalence from ADJ at station 59280 have been compared with the common reference respect to radiosonde equivalence at station 45004. TUT anomalies from MSU at station 59280 and the equivalence at station 45004 have similar variation, and the difference has no significant changes in the 2000s, while the difference between stations 59280 and 45004 shows two significant downward jumps around 2000 and 2010, which are related with the remained discontinuities in ADJ caused by the radiosonde system change in the 2000s (Fig. 9). The comparison demonstrates that the MSU temperature can be used as potential references to the validate homogenization, which is highly valuable for the radiosonde homogenization over China. For the systematic change in Chinese radionde network usually occurs simultaneously, it is hard to find applicative neighbor stations within Chinese network, especially in central China. MSU temperatures with better coverage can provide more useful information on possible breakpoints over China and are helpful to verify or improve the homogenization. Furthermore, the radiosonde equivalence and MSU layer temperatures magnify the difference at individual pressure levels. Temperature differences at individual pressure levels between stations 59280 and 45004 are not so significant as those of the layer mean from MSU and radiosonde equivalence. Therefore, MSU temperatures have better covering and amplifying effects, comparing with the reference at pressure levels from neighbor stations, which contribute to detecting the discontinuity in radiosonde temperature time series. Even so, further research is needed in assigning adjustment from the layer mean to multiple pressure levels.