• Allen, R. J., , and S. C. Sherwood, 2007: Utility of radiosonde wind data in representing climatological variations of tropospheric temperature and baroclinicity in the western tropical pacific. J. Climate, 20 , 52295243.

    • Search Google Scholar
    • Export Citation
  • Arguez, A., 2007: Supplement to State of the Climate in 2006. Bull. Amer. Meteor. Soc., 88 , s1s135.

  • Auligné, T., , and A. McNally, 2006: Adaptive bias correction of satellite data at ECMWF. Proc. ECMWF–NWP-SAF Workshop on Bias Estimation and Correction in Data Assimilation, Reading, United Kingdom, ECMWF, 127–142.

    • Search Google Scholar
    • Export Citation
  • Brohan, P., , J. J. Kennedy, , S. F. B. Tett, , and P. D. Jones, 2006: Uncertainty estimates in regional and global observed temperature changes: A new data set from 1850. J. Geophys. Res., 111 .D12106, doi:10.1029/2005JD006548.

    • Search Google Scholar
    • Export Citation
  • Bronnimann, S., 2003: A historical upper air-data set for the 1939–44 period. Int. J. Climatol., 23 , 769791.

  • Cardinali, C., , S. Pezzuli, , and E. Andersson, 2004: Influence matrix diagnostic of a data assimilation system. Quart. J. Roy. Meteor. Soc., 124 , 17831807.

    • Search Google Scholar
    • Export Citation
  • Charlton, A. J., , and L. M. Polvani, 2007: A new look at stratospheric sudden warmings. Part I: Climatology and modeling benchmarks. J. Climate, 20 , 449469.

    • Search Google Scholar
    • Export Citation
  • Charlton, A. J., and Coauthors, 2007: A new look at stratospheric sudden warmings. Part II: Evaluation of numerical model simulations. J. Climate, 20 , 470488.

    • Search Google Scholar
    • Export Citation
  • Christy, J. R., , W. B. Norris, , W. D. Braswell, , and D. E. Parker, 2003: Error estimates of version 5.0 of MSU–AMSU bulk atmospheric temperatures. J. Atmos. Oceanic Technol., 20 , 613629.

    • Search Google Scholar
    • Export Citation
  • Christy, J. R., , W. B. Norris, , R. W. Spencer, , and J. J. Hnilo, 2007: Tropospheric temperature change since 1979 from tropical radiosonde and satellite measurements. J. Geophys. Res., 112 .D06102, doi:10.1029/2005JD006881.

    • Search Google Scholar
    • Export Citation
  • Compo, G. P., , J. S. Whitacker, , and P. D. Sardeshmukh, 2006: Feasibility of a 100-year reanalysis using only surface pressure data. Bull. Amer. Meteor. Soc., 87 , 175190.

    • Search Google Scholar
    • Export Citation
  • Dee, D. P., 2005: Bias and data assimilation. Quart. J. Roy. Meteor. Soc., 131 , 33233343.

  • Douglass, D. H., , J. R. Christy, , B. D. Pearson, , and S. F. Singer, 2007: A comparison of tropical temperature trends with model predictions. Int. J. Climatol., 28 .doi:10.1002/joc.1651.

    • Search Google Scholar
    • Export Citation
  • Durre, I., , R. S. Vose, , and D. B. Wuertz, 2006: Overview of the Integrated Global Radiosonde Archive. J. Climate, 19 , 5368.

  • Free, M., , and D. J. Seidel, 2005: Causes of differing temperature trends in radiosonde upper air data sets. J. Geophys. Res., 110 .D07101, doi:10.1029/2004JD005481.

    • Search Google Scholar
    • Export Citation
  • Gaffen, D. J., 1996: A digitized metadata set of global upper-air station histories. NOAA Tech. Memo. ERL ARL-211, 38 pp.

  • Gruber, C., , and L. Haimberger, 2008: On the homogeneity of radiosonde wind time series. Meteor. Z., in press.

  • Haimberger, L., 2006: Towards temporally homogeneous evaluations of the observed global atmospheric circulation. Habilitationsschrift (habilitation thesis), University of Vienna, 61 pp. [Available online at http://homepage.univie.ac.at/leopold.haimberger/raobcone1-Dateien/Habilschrift_full.pdf.].

  • Haimberger, L., 2007a: Homogenization of radiosonde temperature time series using innovation statistics. J. Climate, 20 , 13771403.

  • Haimberger, L., 2007b: Homogenization of the global radiosonde temperature and wind dataset. ECMWF Special Project Rep. 2007, 8 pp. [Available online at http://www.ecmwf.int/about/special_projects/haimberger_temp_wind_dataset/report_2007.pdf.].

  • Harris, B. A., , and G. A. Kelly, 2001: A satellite radiance bias correction scheme for data assimilation. Quart. J. Roy. Meteor. Soc., 127 , 14531468.

    • Search Google Scholar
    • Export Citation
  • Held, I. M., , and B. J. Soden, 2006: Robust responses of the hydrological cycle to global warming. J. Climate, 19 , 56865699.

  • Karl, T. R., , S. J. Hassol, , C. D. Miller, , and W. L. Murray, and Eds., 2006: Temperature trends in the lower atmosphere: Steps for understanding and reconciling differences. U.S. Climate Change Science Program and the Subcommittee on Global Change Research Final Rep. Synthesis and Assessment Product 1.1, 180 pp. [Available online at http://www.climatescience.gov/Library/sap/sap1-1/finalreport/sap1-1-final-all.pdf.].

  • Lahiri, S. K., 2003: Resampling Methods for Dependent Data. Springer, 374 pp.

  • Lanzante, J. R., , S. A. Klein, , and D. J. Seidel, 2003: Temporal homogenization of monthly radiosonde temperature data. Part II: Trends, sensitivities, and MSU comparison. J. Climate, 16 , 241262.

    • Search Google Scholar
    • Export Citation
  • Limpasuvan, V., , D. W. J. Thompson, , and D. L. Hartmann, 2004: The life cycle of the Northern Hemisphere sudden stratospheric warmings. J. Climate, 17 , 25842596.

    • Search Google Scholar
    • Export Citation
  • McCarthy, M. P., , H. Titchner, , P. Thorne, , L. Haimberger, , and D. E. Parker, 2008: Assessing bias and uncertainty in the HadAT-adjusted radiosonde climate record. J. Climate, 21 , 817832.

    • Search Google Scholar
    • Export Citation
  • Mears, C. A., , M. C. Schabel, , and F. J. Wentz, 2003: A reanalysis of the MSU channel 2 tropospheric temperature record. J. Climate, 16 , 36503664.

    • Search Google Scholar
    • Export Citation
  • Mears, C. A., , C. E. Forest, , R. W. Spencer, , R. S. Vose, , and R. W. Reynolds, 2006: What is our understanding of the contribution made by observational or methodological uncertainties to the previously reported vertical differences in temperature trends? Temperature trends in the lower atmosphere: Steps for understanding and reconciling differences, U.S. Climate Change Science Program and the Subcommittee on Global Change Research Final Rep. Synthesis and Assessment Product 1.1., 71–88.

  • Onogi, K., and Coauthors, 2007: The JRA-25 reanalysis. J. Meteor. Soc. Japan, 85 , 369432. doi:10.2151/jmsj.85.369.

  • Randel, W. J., , and F. Wu, 2006: Biases in stratospheric and tropospheric temperature trends derived from historical radiosonde data. J. Climate, 19 , 20942104.

    • Search Google Scholar
    • Export Citation
  • Santer, B. D., and Coauthors, 2004: Identification of anthropogenic climate change using a second-generation reanalysis. J. Geophys. Res., 109 .D21104, doi:10.1029/2004JD005075.

    • Search Google Scholar
    • Export Citation
  • Santer, B. D., and Coauthors, 2005: Amplification of surface temperature trends and variability in the tropical atmosphere. Science, 309 , 15511556.

    • Search Google Scholar
    • Export Citation
  • Seidel, D. J., , R. J. Ross, , J. K. Angell, , and G. C. Reid, 2001: Climatological characteristics of the tropical tropopause as revealed by radiosondes. J. Geophys. Res., 106 , 78577878.

    • Search Google Scholar
    • Export Citation
  • Sherwood, S. C., 2007: Simultaneous detection of climate change and observing biases in a network with incomplete sampling. J. Climate, 20 , 40474062.

    • Search Google Scholar
    • Export Citation
  • Sherwood, S. C., , J. Lanzante, , and C. Meyer, 2005: Radiosonde daytime biases and late-20th century warming. Science, 309 , 15561559. doi:10.1126/science.1115640309.

    • Search Google Scholar
    • Export Citation
  • Simmons, A., , S. Uppala, , D. Dee, , and S. Kobayashi, 2007: The ERA interim reanalysis. ECMWF Newsletter, No. 110, ECMWF, Reading, United Kingdom, 25–35.

  • Sperka, S., 2007: Homogeneity adjustments of radiosonde temperature time series using composites of innovations from ERA-40. M.S. thesis, Department of Meterology and Geophysics, University of Vienna, 63 pp. [Available online at http://www.univie.ac.at/img-wien/dipldiss/dipl/DA_Sperka.pdf.].

  • Steiner, A. K., , G. Kirchengast, , M. Borsche, , U. Foelsche, , and T. Schoengassner, 2008: A multi-year comparison of lower stratospheric temperatures from CHAMP radio occultation data with MSU/AMSU records. J. Geophys. Res., 112 .D22110, doi:10.1029/2006JD008283.

    • Search Google Scholar
    • Export Citation
  • Thorne, P. W., , D. E. Parker, , S. F. B. Tett, , P. D. Jones, , M. McCarthy, , H. Coleman, , and P. Brohan, 2005a: Revisiting radiosonde upper air temperatures from 1958 to 2002. J. Geophys. Res., 110 .D18105, doi:10.1029/2004JD005753.

    • Search Google Scholar
    • Export Citation
  • Thorne, P. W., , D. E. Parker, , J. R. Christy, , and C. A. Mears, 2005b: Uncertainties in climate trends: Lessons from upper-air temperature records. Bull. Amer. Meteor. Soc., 86 , 14371442.

    • Search Google Scholar
    • Export Citation
  • Thorne, P. W., and Coauthors, 2007: Tropical vertical temperature trends: A real discrepancy? Geophys. Res. Lett., 34 .L16702, doi:10.1029/2007GL029875.

    • Search Google Scholar
    • Export Citation
  • Titchner, H., , P. W. Thorne, , M. McCarthy, , S. F. B. Tett, , L. Haimberger, , and D. E. Parker, 2008: Critically reassessing tropospheric temperature trends from radiosondes using realistic validation experiments. J. Climate, in press.

    • Search Google Scholar
    • Export Citation
  • Trenberth, K. E., , and L. Smith, 2006: The vertical structure of temperature in the tropics: Different flavors of El Niño. J. Climate, 19 , 49564970.

    • Search Google Scholar
    • Export Citation
  • Trenberth, K. E., and Coauthors, 2007: Observations: Surface and atmospheric climate change. Climate Change 2007: The Physical Science Basis, S. Solomon et al., Eds., Cambridge University Press, 235–336.

    • Search Google Scholar
    • Export Citation
  • Untch, A., , M. Miller, , M. Hortal, , R. Buizza, , and P. A. E. M. Janssen, 2006: Towards a global mesoscale model: The high-resolution system T799L91 and T399L62 EPS. ECMWF Newsletter, No. 108, ECMWF, Reading, United Kingdom, 6–13.

  • Uppala, S. M., 2007: From ERA-15 to ERA-40 and ERA-Interim. Proc. ECMWFGEO Workshop on Atmospheric Reanalysis, Reading, United Kingdom, ECMWF, 17–22.

    • Search Google Scholar
    • Export Citation
  • Uppala, S. M., and Coauthors, 2005: The ERA-40 Re-analysis. Quart. J. Roy. Meteor. Soc., 131 , 29613012.

  • Uppala, S. M., , G. Kelly, , B. K. Park, , P. Kallberg, , and A. Untch, 2006: Experience in estimation of biases in ECMWF reanalyses. Proc. ECMWFNWP–SAF Workshop on Bias Estimation and Correction in Data Assimilation, Reading, United Kingdom, ECMWF, 207–224.

    • Search Google Scholar
    • Export Citation
  • View in gallery

    Decadal temperature trends at radiosonde stations for period 1979–2006, derived (top left) from raw data, (top right) from data homogenized with RAOBCORE, (bottom left) from RSS MSU data and (bottom right) from RICH. Trends are 0000 + 1200 UTC averages, if both time series are available; otherwise, they are derived from either 0000 or 1200 UTC ascents. Here, 400 stations have complete enough time series (at least 26 out of 28 yr of data up to the 30-hPa level). “Cost” quantifies spatial heterogeneity—the smaller the better. Acronyms are explained in section 2.

  • View in gallery

    Time series of monthly mean 0000 UTC temperature anomaly differences with respect to RSS MSU satellite data (a) for the LS and (b) for the TS layer at station Yap (91413, tropical West Pacific). Break dates detected by RAOBCORE are 19820304, 19870904, 19900720, 19951201, and 20021114. Differences are RAOBCORE − RSS (light blue), UAH − RSS (orange), BG − RSS (green), RICH − RSS (pink), HadAT2 − RSS (dark red), and unadjusted (RS) data − RSS (dark blue). Triangles at bottom indicate documented radiosonde instrumentation changes. Corresponding decadal trend differences 1979–2006 (LS) and 1987–2006 (TS) are plotted on the rhs. No TS data are available before 1987.

  • View in gallery

    Same as Fig. 2, but for radiosonde station Bethel (70219, Alaska). Break dates detected by RAOBCORE are 19840404, 19890701, 19920222, 19950724, 19970725, and 19990814: (a), (b) 0000 UTC, (c), (d) 1200 UTC.

  • View in gallery

    Same as Fig. 2, but for radiosonde station Jokioinen (2963, Finland) at 0000 UTC. Break dates detected by RAOBCORE are 19811124, 19860215, and 20050303. Largest differences occur between 1989 and 1993 in January (i.e., in the deepest polar winter).

  • View in gallery

    Zonal mean decadal temperature trends (a) for the LS layer during 1979–2006 and (b) for the TS layer during 1987–2006. Zonal means have been calculated from 10° × 10° box means where at least two radiosonde stations were available. Diamonds are plotted at the center of 10° belts, e.g., 5° for 0°–10°N. Insufficient data available for 90°–70°S and 80°–90°N.

  • View in gallery

    Time series of global mean MSU LS–TS temperature anomaly differences RSS − RAOBCORE (red), UAH − RAOBCORE (orange), ERA-40 BG − RAOBCORE (green), RICH − RAOBCORE (pink), HadAT2 − RAOBCORE (dark red), and unadjusted radiosonde − RAOBCORE (dark blue). Corresponding trend differences are shown on the rhs. All global means are from data subsampled at radiosonde stations. Note that in contrast to the station intercomparisons, the anomaly differences are now with respect to RAOBCORE. RAOBCORE trends for the LS and TS layers are −0.4 K (decade)−1 and −0.04 K (decade)−1, respectively. The large variance of the HadAT2 − RAOBCORE difference series is caused by the much coarser spatial sampling of the HadAT2 dataset south of the equator.

  • View in gallery

    Same as Fig. 6, but for the tropics (20°S–20°N). The figure has been derived from 42 stations in the tropics with 26 out of 28 yr of data since 1979. Difference of dark blue curve in the 1980s minus the 2000s yields RAOBCORE bias estimates (approx 1 K in the LS; approx 0.6 K in the TS). RAOBCORE trends for the LS and TS layers are −0.34 K (decade)−1 and −0.02 K (decade)−1, respectively.

  • View in gallery

    Decadal trend differences with respect to RAOBCORE for the (top) LS layer and (bottom) TS layer. Differences are calculated (a) for the tropics during nighttime [0000 UTC between 90°W and 90°E, 1200 UTC between 90°E and 90°W, RAOBCORE trends −0.32 (−0.06) K (decade)−1 for the LS (TS) layer], (b) for the tropics during daytime [RAOBCORE trends −0.36 (−0.02) K (decade)−1 for the LS (TS) layer], (c) for the northern extratropics (20°–90°N) daily means [RAOBCORE trends −0.37 (0.07) K for the LS (TS) layer], (d) for the southern extratropics (20°–90°S) daily means [RAOBCORE trends −0.48 (−0.13) K (decade)−1 for the LS (TS) layer].

  • View in gallery

    Mean effect of radiosonde temperature adjustments in the tropics (20°N–20°S) for the (a) LS and (b) TS layer. The figure in the title is the number of radiosonde stations that have been adjusted. Thick curves show effect of large breaks (>0.5 K for the LS layer; >0.3 K for the TS layer), and thin curves show effect of breaks smaller than these thresholds. Blue curves show effect of adjustments estimated by RAOBCORE. Red curves are effect of adjustments calculated from RSS-unadjusted radiosonde differences using the same time intervals as RAOBCORE. Practically the same curve as with RSS LS data can be obtained with UAH LS data (not shown). Pink curves are calculated from adjustments estimated with RICH.

  • View in gallery

    Same as Fig. 7, but with anomaly differences with respect to RAOBCORE, version 1.3, where the BG has not been corrected before radiosonde adjustment at all, not even before 1987. HadAT2 series is omitted here for clarity and because it is already plotted in Fig. 7.

  • View in gallery

    Vertical temperature trend profiles (a) for the tropics (20°S–20°N) and (b) for the global mean. Thick solid curve is standard RICH estimate using standard eight reference stations. Thick dashed–double dotted curve is RICH estimate using 30 reference stations. Thin solid curve is RAOBCORE, version 1.4, estimate, thin dashed is RAOBCORE, version 1.3, and dotted is from unadjusted radiosonde data. HadAT2 profiles (thin dashed–dotted) are estimated from less available radiosondes and are included for reference. Corresponding surface temperature trends from HadCRUT, version 3.0 are denoted with x symbols.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 78 78 19
PDF Downloads 45 45 6

Toward Elimination of the Warm Bias in Historic Radiosonde Temperature Records—Some New Results from a Comprehensive Intercomparison of Upper-Air Data

View More View Less
  • 1 Department of Meteorology and Geophysics, University of Vienna, Vienna, Austria
© Get Permissions
Full access

Abstract

The apparent cooling trend in observed global mean temperature series from radiosonde records relative to Microwave Sounding Unit (MSU) radiances has been a long-standing problem in upper-air climatology. It is very likely caused by a warm bias of radiosonde temperatures in the 1980s, which has been reduced over time with better instrumentation and correction software. The warm bias in the MSU-equivalent lower stratospheric (LS) layer is estimated as 0.6 ± 0.3 K in the global mean and as 1.0 ± 0.3 K in the tropical (20°S–20°N) mean. These estimates are based on comparisons of unadjusted radiosonde data, not only with MSU data but also with background forecast (BG) temperature time series from the 40-yr European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA-40) and with two new homogenized radiosonde datasets. One of the radiosonde datasets [Radiosonde Observation Correction using Reanalyses (RAOBCORE) version 1.4] employs the BG as reference for homogenization, which is not strictly independent of MSU data. The second radiosonde dataset uses the dates of the breakpoints detected by RAOBCORE as metadata for homogenization. However, it relies only on homogeneous segments of neighboring radiosonde data for break-size estimation. Therefore, adjustments are independent of satellite data.

Both of the new adjusted radiosonde time series are in better agreement with satellite data than comparable published radiosonde datasets, not only for zonal means but also at most single stations. A robust warming maximum of 0.2–0.3K (10 yr)−1 for the 1979–2006 period in the tropical upper troposphere could be found in both homogenized radiosonde datasets. The maximum is consistent with mean temperatures of a thick layer in the upper troposphere and upper stratosphere (TS), derived from M3U3 radiances. Inferred from these results is that it is possible to detect and remove most of the mean warm bias from the radiosonde records, and thus most of the trend discrepancy compared to MSU LS and TS temperature products.

The comprehensive intercomparison also suggests that the BG is temporally quite homogeneous after 1986. Only in the early 1980s could some inhomogeneities in the BG be detected and quantified.

* Current affiliation: European Centre for Medium-Range Weather Forecasts, Reading, United Kingdom

Corresponding author address: Leopold Haimberger, Department of Meteorology and Geophysics, University of Vienna, Althanstrasse 14, A-1090 Vienna, Austria. Email: leopold.haimberger@univie.ac.at

Abstract

The apparent cooling trend in observed global mean temperature series from radiosonde records relative to Microwave Sounding Unit (MSU) radiances has been a long-standing problem in upper-air climatology. It is very likely caused by a warm bias of radiosonde temperatures in the 1980s, which has been reduced over time with better instrumentation and correction software. The warm bias in the MSU-equivalent lower stratospheric (LS) layer is estimated as 0.6 ± 0.3 K in the global mean and as 1.0 ± 0.3 K in the tropical (20°S–20°N) mean. These estimates are based on comparisons of unadjusted radiosonde data, not only with MSU data but also with background forecast (BG) temperature time series from the 40-yr European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA-40) and with two new homogenized radiosonde datasets. One of the radiosonde datasets [Radiosonde Observation Correction using Reanalyses (RAOBCORE) version 1.4] employs the BG as reference for homogenization, which is not strictly independent of MSU data. The second radiosonde dataset uses the dates of the breakpoints detected by RAOBCORE as metadata for homogenization. However, it relies only on homogeneous segments of neighboring radiosonde data for break-size estimation. Therefore, adjustments are independent of satellite data.

Both of the new adjusted radiosonde time series are in better agreement with satellite data than comparable published radiosonde datasets, not only for zonal means but also at most single stations. A robust warming maximum of 0.2–0.3K (10 yr)−1 for the 1979–2006 period in the tropical upper troposphere could be found in both homogenized radiosonde datasets. The maximum is consistent with mean temperatures of a thick layer in the upper troposphere and upper stratosphere (TS), derived from M3U3 radiances. Inferred from these results is that it is possible to detect and remove most of the mean warm bias from the radiosonde records, and thus most of the trend discrepancy compared to MSU LS and TS temperature products.

The comprehensive intercomparison also suggests that the BG is temporally quite homogeneous after 1986. Only in the early 1980s could some inhomogeneities in the BG be detected and quantified.

* Current affiliation: European Centre for Medium-Range Weather Forecasts, Reading, United Kingdom

Corresponding author address: Leopold Haimberger, Department of Meteorology and Geophysics, University of Vienna, Althanstrasse 14, A-1090 Vienna, Austria. Email: leopold.haimberger@univie.ac.at

1. Introduction

Despite much research in upper-air observations there is still sizeable discrepancy between layer mean atmospheric temperatures derived from radiances of the Microwave Sounding Unit (MSU) and MSU-equivalent temperatures calculated from radiosonde data (see, e.g., Santer et al. 2005; Karl et al. 2006; Trenberth et al. 2007). Mears et al. (2006) suspect that especially the lower stratospheric time series from radiosondes are biased and Sherwood et al. (2005) and Randel and Wu (2006) have described pervasive daytime biases in the tropics. Allen and Sherwood (2007) have further shown that the apparent cooling in tropical radiosonde temperatures is not consistent with wind shear changes in this region. However, all this evidence has been fragmentary so far and the radiosonde biases have yet to be removed in a comprehensive manner. This paper reports about new radiosonde homogenization efforts to remove these biases. It tries to give a firmer answer on the magnitude of the raw radiosonde temperature mean warm bias and on the existence of important upper-air climate features in radiosonde data, such as the tropical upper tropospheric warming maximum (Santer et al. 2005). For this purpose, a detailed intercomparison between satellite data, radiosonde data, and 40-yr European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA-40) products has been undertaken.

With the advent of both the Integrated Global Radiosonde Archive (IGRA; Durre et al. 2006) and the so-called ERA-40 analysis feedback dataset (Uppala et al. 2005), amended with the operational ECMWF analysis feedback from 2001 onward, the density of radiosonde datasets has recently been improved. Both datasets reach back to at least 1958 and are inputs to the Radiosonde Observation Correction using Reanalyses (RAOBCORE) dataset (Haimberger 2007a), which contains more than 1000 adjusted radiosonde temperature time series. RAOBCORE uses time series of the ERA-40 background forecasts (BG) as reference for homogenization. The RAOBCORE adjustments make the radiosonde data spatially more consistent, which is a necessary, though not sufficient condition for homogeneity. They also remove most of the artificial trends in the diurnal temperature cycle.

Nevertheless, the uncertainty estimates for global mean trends from early versions of this dataset are on the order of 0.3 K (decade)−1 for the MSU lower stratospheric (LS) layer, because some periods of the BG time series have been found questionable (Haimberger 2007a; Uppala et al. 2005; Trenberth and Smith 2006; Karl et al. 2006). The uncertainty is also caused by the unknown degree of dependence of RAOBCORE adjustments on MSU radiances, because these are input into the ERA-40 and the operational analysis systems and thus influence the BG. Here, we introduce a new automatic homogenization method called Radiosonde Innovation Composite Homogenization (RICH). It constructs the reference series for homogenization from composites of neighboring radiosonde station temperature anomalies. Therefore, its break-size estimates are independent of satellite data. RICH uses the more than 6000 breakpoint dates that have been detected by RAOBCORE as metadata to avoid the use of inhomogeneous pieces of neighboring reference time series. The generation of these metadata may have profited from the use of MSU radiances in ERA-40, and thus these may have influenced the selection of reference stations. Because only radiosonde data are used in the samples for break-size estimation, the RICH method is nevertheless considered independent. A preliminary description of RICH has been provided by Sperka (2007). More details are given in section 2 below.

To quantify any progress made compared to the state-of-the-art, the Hadley Centre Atmospheric Temperature (HadAT2; Thorne et al. 2005; 676 stations) dataset is included in this intercomparison. While we have also applied the comparison procedure to the so-called radiosonde atmospheric temperature products for assessing climate (RATPAC; Free and Seidel 2005; 87 stations) dataset, we show no results here because we want to demonstrate the overall robustness of the homogeneity adjustment approaches when applied to all available radiosonde data, not just to a relatively small subset. A global belt–mean comparison between RAOBCORE and RATPAC is already published (Arguez 2007).

From 1979 onward, the radiosonde data can be compared with MSU brightness temperatures representative for thick vertical layers. We used Remote Sensing Systems (RSS) version 3.0 (Mears et al. 2003) and University of Alabama at Huntsville (UAH) version 5.1 (Christy et al. 2003) brightness temperatures representative for the lower stratosphere [from MSU4 and Advanced Microwave Sounding Unit (AMSU9)]. Because we concentrate on the upper troposphere and stratosphere, the mid- and lower tropospheric channels have not been considered in this paper (results for these are provided on the authors’ Web site; http://www.unvie.ac.at/theoret-met/research/raobcore). Instead, we included data from the MSU3 and AMSU7 channels in the intercomparison, which are not sensitive to the surface. These channels sample a very thick layer (peak sensitivity at approximately 250 hPa in the upper troposphere and lower stratosphere, which is referred to as the TS layer, and the corresponding temperature product is provided by RSS. The unusual behavior of the MSU3 radiances on the National Oceanic and Atmospheric Administration NOAA-6 and NOAA-9 satellites and a data gap in 1986, which is very difficult to handle (Mears et al. 2006; Uppala et al. 2005) are the reasons why RSS TS products are available only from 1987 onward. Yet, we consider radiances from this channel as important for the following reasons: (i) they are routinely assimilated in climate data assimilation systems, where they have substantial influence on the analyzed upper-tropospheric temperatures; (ii) the much-debated upper-tropospheric warming in the tropics has its maximum near the level of peak sensitivity of the TS product; (iii) because it is an independent channel, it is certainly valuable for consistency checks; and (iv) the TS product has not been systematically compared with radiosonde data yet. While our main focus is on detection of biases in radiosonde data, there may be residual biases in the LS and TS satellite products as well. Recent comparisons with GPS radio occultation data (Steiner et al. 2008) indicate that temperature products derived from AMSU data may be less reliable than expected.

In the next two sections, the input data and the comparison method are described in some detail. In section 4, a few intercomparisons for the MSU LS and TS equivalent layers at individual radiosonde sites are presented. In section 5, global-belt mean comparisons are discussed and it is shown how the discrepancy between unadjusted radiosondes and satellite data can be explained as the composite effect of relatively well-detectable artificial shifts in the radiosonde time series. Also, the effect of some difficulties of the BG time series in the early 1980s could be quantified. Implications of these results are discussed in section 6 and in the conclusions.

2. Input data

The following input datasets have been used for this intercomparison:

  • RSS MSU, version 3.0, data (Mears et al. 2003) on the standard 2.5° × 2.5° latitude–longitude grid at LS and TS layers.
  • UAH MSU, version 5.1, data (Christy et al. 2003) on the standard 2.5° × 2.5° latitude–longitude grid at the LS layer. The RSS and UAH temperature datasets have been calculated from the same raw radiance datasets, though with quite different analysis methods. Both have monthly time resolution. The main uncertainties arise from the connection of data from different satellite platforms that often have rather short overlap time and from the correction of drift effects. For comparison with radiosonde data, the MSU data have been subsampled to the radiosonde station locations (see next section).
  • ERA-40 and operational ECMWF BG on a 1.25° × 1.25° latitude–longitude grid and interpolated to radiosonde stations. This dataset has 6-hourly resolution but only daily 0000 and 1200 UTC data are used. The BG is the product of a comprehensive data assimilation procedure that uses most available observations, including radiosonde data, surface data, and satellite radiances. The data for ERA-40 have been assimilated with a three-dimensional variational [3D-VAR; first guess at appropriate time (FGAT); Uppala et al. 2005] system, which is beneficial for assimilation of satellite radiances, especially if there are orbital drifts. Because of the large differences in the assimilation system compared to UAH and RSS, it is by no means clear that the resulting BG temperature time series are consistent with the RSS and UAH satellite retrievals.

BG data from ERA-40 have been used for the period 1958–2000. From 2001 onward, operational ECMWF background forecasts have been used. These are 12-h forecasts that are part of the four-dimensional variational (4D-VAR) assimilation system with 12-h cycling. The fraction of information content coming from radiosondes in this recent period is relatively small (Cardinali et al. 2004). However, the influence of radiosondes on biases may be larger because the bias-correction procedure applied in ERA-40 and in ECMWF operations until early 2006 (Harris and Kelly 2001) heavily relied on comparison with a subset of high-quality radiosondes.

The BG is included in this intercomparison not only because it is used as reference for the RAOBCORE homogenization but also to show that its temporal homogeneity is competitive with pure satellite or radiosonde datasets over long time intervals.

RAOBCORE works with daily data and uses the BG as reference for automatic homogenization of the available radiosonde temperature records. The homogenized radiosonde data used in this paper (version 1.4) have been generated as described in Haimberger (2007a), with the exception that the modification of the global-mean BG time series has been applied only before 1987. This is a compromise between the two versions used by Haimberger (2007a), where the BG was modified either throughout the whole period (1958–2005) or was not modified at all. The reasons for this choice are explained in section 5 below.

  • The updated HadAT2 dataset (Thorne et al. 2005; McCarthy et al. 2008). It is the most complete homogeneity-adjusted pure radiosonde dataset published so far and consists of 676 station series, ranging from 1958 to the present. Daytime and nighttime soundings are not kept separate. The input data resolution is monthly, and the adjustments are provided with seasonal resolution.
  • Radiosonde data adjusted with RICH. This dataset uses the same raw radiosonde data as input as does RAOBCORE; that is, it also consists of more than 1000 daily station records and it works on 0000 and 1200 UTC time series separately. It uses composites of homogeneous pieces of temperature time series from neighboring radiosonde stations as reference for break-size estimation. In this regard, the method is similar to the HadAT2 approach. Because only radiosonde data are used, the RICH break-size estimation is independent of satellite data.

However, RICH does not detect the breakpoints themselves but uses the breakpoint dates obtained from RAOBCORE as metadata. These breakpoint dates determine homogeneous pieces of radiosonde time series, which are used for break-size estimation. The time series of 3 month after a breakpoint until 3 month before the next breakpoint are considered homogeneous. The 3-month offset alleviates the effect of possible inaccuracies in the detected break dates. MSU data can influence RICH estimates only via the breakpoint dates provided by RAOBCORE.

The reference series used for estimation of the break size are built strictly from the homogeneous pieces of the temperature time series at neighboring radiosonde sites. These pieces can be rather short and such a restrictive criterion could be applied only because daily data have been used, which allowed break-size estimation for time series as short as 130 days. This is much shorter than has been possible in the HadAT2 adjustment process (Thorne et al. 2005; McCarthy et al. 2008), which used monthly and seasonal data.

In contrast to the preliminary version of RICH described by Sperka (2007), who used large composites (50–200 stations), the present implementation of RICH uses no more than eight neighboring radiosondes for building a reference series. This helps to keep the distance between tested stations and reference stations low. It also reduces the risk of including a station with undetected breaks in the composite. Most importantly, however, it reduces the risk of introducing a cooling bias when constructing a reference for tropical stations near the tropopause level (approximately 100 hPa; Seidel et al. 2001). While the 100-hPa level is near or below the tropopause in the inner tropics, it lies in the cooling stratosphere in the extratropics. Larger composites have the advantage that a missed breakpoint in one reference series is less detrimental than for small composites, and they generally lead to spatially smoother trend patterns. A sensitivity experiment with RICH using larger composites is described below.

The reference series are weighted means of the anomalies from neighboring time series, where the weights decrease exponentially with distance from the tested stations and where the distance in the meridional direction is weighted higher than the distance in the zonal direction. Despite the small number of reference stations and the sometimes short time intervals available for break-size estimation, the obtained spatial trend consistency is only slightly worse than obtained with RAOBCORE (see Fig. 1). A further increase of spatial consistency could have been easily achieved by using more reference stations, but this does not necessarily lead to less-biased trends.

We note that Sperka (2007) has used not only anomalies of neighboring time series but also ERA-40 background departures (innovations) of neighboring time series for building reference series. In the present paper, however, only anomalies are used to preserve independence of satellite data.

3. Comparison method

The comparison of pressure-level data at radiosonde stations with gridded-layer mean satellite temperatures is nontrivial, and one must ensure that differences between these datasets are genuine and not caused by sampling errors, interpolation errors, or other artifacts (Free and Seidel 2005). Therefore, we outline our comparison strategy in the following subsections.

a. Construction of the MSU equivalents

The radiosonde data as well as the ERA-40–ECMWF operational products are available on standard pressure levels (10, 20, 30, 50, 70, 100, 150, 200, 250, 300, 400, 500, 700, and 850 hPa) and are valid for the station location. From these data, MSU equivalent-layer means can be calculated. The necessary weighting functions for the LS layer are the same as those used by Thorne et al. (2005). For the TS layer, weighting functions have been provided by C. Mears (2006, personal communication).

b. Construction of anomalies

MSU equivalent time series have been calculated from radiosonde data and BG time series separately for 0000 and 1200 UTC. At least 10 values per month have been necessary to yield a monthly mean value valid for 0000 and 1200 UTC, respectively. The RSS and UAH data have been taken as available and can be interpreted as monthly means of daily means with no distinction between 0000 and 1200 UTC.

The satellite products are available as anomalies with respect to the 1979–98 climatology. The 1979–98 climatologies of the BG and radiosonde data have also been subtracted from each series to create anomalies. Much care is needed for this task, especially if the time series contain large and highly transient climatological signals such as stratospheric sudden warmings (SSW) and if values are missing. The large amplitude of these warmings (up to 15 K in the anomalies of LS MSU equivalents; Limpasuvan et al. 2004; Charlton and Polvani 2007; Charlton et al. 2007) has the potential to substantially influence the climatologies. Therefore, climatologies and anomalies have been calculated only if at least 15 out of 20 yr have been available and only for those months that have been available in all (MSU, BG, and radiosonde) datasets noted above. This procedure guarantees that differences between datasets at individual stations and in the global means cannot be related to missing monthly data.

An exception has been made only for the comparisons at the three example stations in Figs. 2 –4, in which we calculated anomalies for the shorter 1996–2006 period to make the seasonal cycle of the radiation error more visible.

c. Horizontal interpolation and sampling

This paper reports about individual station intercomparisons as well as global-scale intercomparisons on a monthly time scale. The radiosonde data and the BG time series are valid at the observation locations. The nearest gridpoint with MSU data, which are available on a 2.5° × 2.5° grid, has been used for station-by-station intercomparisons. This should be accurate enough for the large-scale features considered.

Only radiosonde time series with less than 24 months of missing data in the 28-yr period 1979–2006 have been included in the global and tropical belt–mean intercomparisons. The 24-month threshold is necessary for a stable trend calculation. Figure 1 shows LS-equivalent trends of all the radiosonde stations that fulfill this criterion.

One problem in comparing radiosonde and MSU data is the uneven distribution of radiosonde data (Free and Seidel 2005). For zonal and global means, we used the following sampling strategy:

  • For each 10° × 10° grid box, a gridbox mean temperature has been calculated from the available long radiosonde time series at 0000 and 1200 UTC. For some grid boxes over Europe and China, this is an average over 20 radiosonde time series; for others, it is only one radiosonde time series. No Indian stations have been used because their temperature series are rather inaccurate (Lanzante et al. 2003), at least in the 1980s and 1990s.
  • For each 10° × 10° grid box, an average MSU brightness temperature has been calculated from those grid points that have been nearest neighbors to a radiosonde station.
  • Each 10° latitudinal belt has 36 such grid boxes with values at 0000 and 1200 UTC. These values have been averaged to yield a daily mean zonal belt mean, if at least 2 out of 36 boxes were available. The zonal belt means have been averaged to yield global and tropical means.

4. Station intercomparisons—Three examples

The noise level of different time series from MSU-equivalent radiosonde temperatures and MSU satellite retrievals is small enough to allow sensible comparisons at individual stations. To illustrate the value of such an intercomparison, three representative examples have been chosen.

Figure 2 shows differences between RAOBCORE − RSS, RICH − RSS, UAH − RSS, BG − RSS and Unadjusted (RS) radiosondes − RSS of the LS- and TS-equivalent-layer mean temperature anomaly time series at station Yap (91413, tropical West Pacific). The time series have been filtered with a 3-month running mean. It has been decided to plot differences of anomalies, because for some stations the anomalies at high latitudes are very large (15 K during SSWs). The much smaller differences between the datasets would be hard to spot in a normal anomaly plot. The choice of reference is arbitrary, so the fact that the time series in Fig. 2a show differences with respect to the RSS MSU dataset does not imply that this dataset is the most accurate one.

One can see that the BG time series, UAH data, and RSS data are in good agreement in the LS layer, compared to the large differences between unadjusted and adjusted radiosonde datasets. The break in 1996 is just one example of the large inhomogeneities found in several radiosonde records. Similar large breaks can be found at other tropical stations (e.g., 91348, 91592, and 91958). The breaks are so large that they can safely be attributed to the radiosonde time series. The treatment of MSU radiances in RSS, UAH, and ERA-40 is sufficiently different, and their uncertainties are sufficiently small (see Thorne et al. 2007 and references therein), to exclude the possibility that these three datasets are affected by a jump of this size in MSU radiances.

The adjusted radiosonde temperature time series shows reasonable agreement with the satellite and the BG time series. RAOBCORE delivers the largest adjustments at Yap. RICH yields slightly smaller adjustments, using the same break dates but using neighboring radiosondes. HadAT2 shows weaker adjustments, and it seems evident that the break date in 1996 is not exactly hit.

The right panels of Fig. 2a show the differences in trends for 1979–2006, also with respect to RSS data. For the LS layer, the BG, RAOBCORE, and RICH trend estimates are practically indistinguishable from the satellite trends, given a typical trend difference sampling uncertainty of 0.1 K (decade)−1 at such stations. The HadAT2 adjustments left some residual cooling compared to satellite data. Yap has reported only at 0000 UTC most of the time. Thus, the HadAT2 estimates are valid for 0000 UTC as well.

Figure 2b shows a detailed intercomparison of MSU3–(TS)-equivalent time series. Please note that the temperature scale is different in this plot. The results from Fig. 2a are basically corroborated by data from the TS-equivalent layer. However, there remains a cooling bias in all radiosonde records compared to the RSS products in this layer, which needs further study.

Although we focus on the tropics, we have included two stations in polar regions to illustrate the power of the present analysis for detecting data discrepancies. One example is Bethel (70219, Alaska), which has been studied in some detail already by Haimberger (2007a). The comparison with MSU-equivalent anomaly time series complements the description of this station, at least for the satellite period. Note that the MSU time series and HadAT2 series are daily means, while the unadjusted RS, RAOBCORE, and RICH series are different for 0000 and 1200 UTC. Three types of radiosondes with very different radiation errors were in use at Bethel (VIZ before 1989, space data radiosondes between 1989 and 1995, Vaisala RS80 from 1995 onward). In Fig. 3, the anomalies have been calculated over the Vaisala period (1996–2006), and one can see good agreement among all datasets after 1995. During the space data period, extremely large daytime biases due to radiation errors have occurred, and these are very well detectable in both the LS and TS layers. The strong annual cycle in the radiation errors is typical for stations near or poleward of the polar circle. There is also a substantial nighttime (1200 UTC) cool bias in the unadjusted radiosonde data during this period. Before 1989, when VIZ-A radiosondes were in use, moderate radiation errors can be seen. It is clear that seasonally constant adjustments cannot remove the strong annual cycle in the radiation error, but at least the annual mean bias could be removed reasonably well by all three radiosonde homogenization methods.

Time series from stations using Vaisala radiosondes for a long time are considered as quite reliable by many researchers. Therefore, we investigated the Finnish station Jokioinen (2935) to check if there are noticeable undetected breaks in Vaisala radiosonde time series. Figure 4 shows that the agreement with satellite data is quite good at this station. There is, however, an interesting period (1989–92), where there is a strong annual cycle in the differences, not only between RSS–MSU data and radiosonde data but also between RSS and UAH and even more between RSS and ERA-40 BG. The simplest explanation would be that the operational radiation correction was erroneous before 1993. In 1993, a new radiation correction was been introduced by most stations launching Vaisala radiosondes. Because Vaisala radiosondes are widespread in Europe, it seems possible that even the BG has been affected by this error. It is, however, surprising that this radiosonde − RSS difference is largest in the deepest polar winter and also that the RSS − UAH difference shows a noticeable annual cycle during this period. We therefore cannot exclude the possibility that the RSS dataset is in error during this period as well.

In addition to the annual cycle, there seems to be an annual mean bias change in 1993 compared to RSS, which has not been adjusted by RAOBCORE and HadAT2. If this bias change is representative for many stations using Vaisala radiosondes at this time, it may be the reason for the relatively strong cooling trend compared to RSS over Europe in Fig. 1. More detailed analysis will be necessary to trace the cause of these deviations.

Plots of difference series as shown in Figs. 2 –4 are available online (http://www.univie.ac.at/theoret-met/research/RAOBCORE) for more than 600 radiosonde stations for which climatologies could be calculated. They allow a systematic station-by-station evaluation of the discrepancies among the different datasets and are available for tropospheric layers as well. We consider it very important to show that the homogenization procedures presented here improve most single-station records. The plots leave little doubt that most large breaks are attributable to radiosondes and that the signal-to-noise ratio of difference series between monthly MSU data and radiosonde data is surprisingly good in general.

5. Comparison of global and tropical layer–mean temperatures

In this section, the various temperature time series for 1979–2006 are compared using global maps of trends at individual stations and global-belt mean time series. From Fig. 1 one can see the improved spatial trend homogeneity in the LS layer achieved with RAOBCORE and RICH compared to the unadjusted radiosondes. The spatial trend homogeneity is an important measure for the overall robustness of the adjustment methods, because a few badly adjusted records are sufficient to make the spatial heterogeneity measure (“Cost” in the figure titles, defined in Haimberger 2007a) worse than that of the unadjusted radiosondes. The trends of both adjusted radiosonde datasets in Fig. 1 show better consistency than the unadjusted dataset. Because the individual-station trends denoted by the bullets agree much better with the RSS trends (and UAH trends, not shown), it can also be concluded that the adjusted data fit much better to MSU data than the unadjusted data. This confirms the results found for the three individual stations. However, there are still interesting differences: for example, the cooling in northern midlatitudes, which is weaker in RSS than in both radiosonde datasets and UAH. The patterns of UAH trends (not shown) are as smooth as RSS, but they show consistently more cooling, as may be seen from Fig. 5.

The zonal-mean trends plotted for the LS (Fig. 5a; 1979–2006) and TS (Fig. 5b; 1987–2006) layers show the much better agreement of RAOBCORE- and RICH-adjusted data with satellite data compared to unadjusted radiosondes. Only the data-sparse southern extratropics are an exception. HadAT2 zonal-mean trends are plotted as well for reference. Because of the stringent data-availability requirements imposed here on the radiosonde time series, the zonal means from HadAT2 look incomplete in the LS plot.

While the maps give an impression of the spatial homogeneity of the data, the overall temporal homogeneity is best examined in terms of anomaly difference time series of large-scale means. We concentrate on the tropics (20°S–20°N) and on global means. Again, we do not show anomaly time series, because their shape is well known, and plots of global anomalies including RAOBCORE and HadAT2 have been recently published (Arguez 2007).

In contrast to the station intercomparisons in Figs. 2 –4, where we plotted the anomaly differences with respect to RSS data, we now plot anomaly differences with respect to the homogenized RAOBCORE, version 1.4, radiosonde time series. In such plots, the dark blue lines in Figs. 6, 7, which show unadjusted radiosondes − RAOBCORE differences, can be interpreted as the “anomalies” of the adjustments applied by RAOBCORE. If we take the difference between the adjustment anomalies in the early 1980s and the adjustment anomalies in 2006, we get the RAOBCORE estimate for the mean warm bias of radiosonde temperatures in the lower stratosphere in the early 1980s. It is about 0.6 K in the global mean LS layer and 1 K in the tropical mean LS layer.

The other global anomaly differences in Fig. 6 show the noticeable warming of the RSS MSU data and BG time series in LS and TS layers compared to UAH and adjusted radiosonde data. However, the LS and TS trends of RAOBCORE, RICH, satellite data, and BG differ by only 0.1 K (decade)−1 (Fig. 6, rhs), whereas the trend difference between unadjusted radiosonde and satellite data is 0.4 K (decade)−1 (see also Table 1). The global mean trends from HadAT2 are approximately 0.2 K (decade)−1 lower than RSS.

When turning to the tropics, one can see that the differences between RAOBCORE data, RICH data, MSU data, and BG are less than 0.3 K throughout the TS series in Fig. 7, and with few exceptions also in the LS series. The jump of the BG in 2006 compared to the other datasets is such an exception, which has been caused by the profound change in the ECMWF operational assimilation system in February 2006 (Untch et al. 2006). In the RAOBCORE versions discussed here, no attempt to adjust this shift in the BG adjustment is made. To avoid possible negative effects of this shift, no data after January 2006 are used in RAOBCORE for break detection, and the most recent breaks (after mid-2005) are left unadjusted.

The dark blue difference curves between unadjusted and adjusted radiosonde time series show large and systematic deviations, for the TS and even more so for the LS layers. The difference in the tropics is 1 K in the LS layer and 0.6 K in the TS layer. Not surprisingly, the trend difference between unadjusted and RAOBCORE–RICH adjusted radiosonde data in the tropical LS layer is more than 0.4 K (decade)−1 (Fig. 7, rhs).

It is interesting to look at tropical daytime and nighttime trend differences separately, as shown in Figs. 8a,b. Please note that we find cooling biases in the unadjusted tropical data in both the daytime and the nighttime series. While the nighttime biases are weaker, we find no sign of compensation between daytime and nighttime biases, at least in the tropical mean LS and TS layers. RAOBCORE and RICH consistently reduce the cooling bias at both daytime and nighttime. Compared to MSU trend estimates, the adjusted daytime trends from RICH and RAOBCORE show only slightly [less than 0.1 K (decade)−1] more cooling than nighttime trends.

For completeness, Figs. 8c,d show 24-h mean trends (i.e., no separation between daytime and nighttime) for the extratropics. In the northern extratropics (Fig. 8c) the RAOBCORE–RICH trends are again much warmer than the unadjusted trends and lie just between RSS and UAH in the LS layer. There is also reduction of the cooling bias with respect to RSS in the TS layer. In the southern extratropics, RAOBCORE trends appear fairly consistent with the satellite data, while RICH appears to have a problem due to the scarcity of neighboring radiosonde stations. An extra effort is needed to fix this problem in RICH. The results from Fig. 8 are consistent with the zonal means shown in Fig. 5.

The time series of differences alone do not imply that the satellite data and BG are right and the radiosondes are biased. One could still argue that the differences have been generated by many too-liberal homogeneity adjustments that simply draw the radiosonde time series to the BG time series.

We therefore quantified the systematic effect of adjustments of large breaks detected in radiosonde time series such as in Fig. 2 to see if they can explain the tropical mean trend discrepancies. Figure 9 shows the composite effect of adjustments of large (>0.5 K in the LS, >0.3 K in the TS layer) breaks at individual radiosonde sites. Breaks of such size are statistically highly significant, are well detectable, and can be attributed to radiosonde records because differences between MSU datasets and BG are much smaller, even on small spatial and temporal scales.

For this figure, the break sizes at individual radiosonde sites have been estimated not only by RAOBCORE, where the BG used the reference, but also by using RSS or UAH MSU data as well as neighbor composites as references (i.e., by using RICH). The cumulative effects of the resulting LS-equivalent adjustments gained from these three different reference datasets agree very well with each other and also agree quite well in size and shape with the adjustments in Fig. 7.

We infer from this result that the apparent cooling of radiosonde time series can be explained by the cumulative effect of large shifts. This is especially true if there are known changes in instrumentation, which is the case for most large breaks.

It is remarkable that we can reproduce the homogeneity adjustments obtained to such a large extent with RAOBCORE and also with RICH. The hypothesis formulated by Thorne et al. (2005)—that it is possible to homogenize the radiosonde temperatures with composites of neighboring radiosonde data—now appears better justified than ever.

The systematic effect of smaller breaks (the dotted curves in Fig. 9) is limited and almost disappears if we reduce the threshold for “large” to 0.3 (0.2) K in the LS (TS) layer. Small breaks have a more systematic effect when using the RSS dataset as reference, which is consistent with the warming of RSS data compared to the other datasets and still visible in Figs. 6 –8.

This discrepancy shows that the effect of small breaks is not negligible when discussing trends, but neither can it alter the overall conclusion that a very large fraction (at least 75%) of the differences between radiosonde and satellite trends in the LS layer is attributable to problems with radiosondes in the 1980s. The trends from the radiosonde time series adjusted with RICH and RAOBCORE in Fig. 7 are still slightly more toward cooling than those from RSS satellite time series. The reason may be either that a few breaks in the radiosonde time series are still undetected or that the satellite trend estimates are biased high, or both. The differences between RSS, RAOBCORE, and RICH are at most 0.3 K in the LS layer in Figs. 9, 7.

The same figures show qualitatively similar results for the TS layer. There is, however, only one MSU dataset (RSS) available from 1987 onward. Trends from RSS TS data are more toward warming than the BG and radiosonde data. Small shifts between the RSS and the BG are noticeable in 1989, 1995, and 2003. The shift in 1989 coincides with a change of “assimilation streams” and with a significant increase of satellite data available for assimilation in ERA-40. The shift in 1995 coincides with a change from NOAA-11 to NOAA-14, and the shift in 2003 occurred at a time when substantial changes in the operational ECMWF system became operational. These shifts deserve closer examination at a later stage so that they can be safely attributed to a specific data source. Therefore, the uncertainty involved in the satellite data, the BG, and the adjusted radiosonde data is only slightly smaller than for the LS layer, although the anomalies and the trend signal are weaker. The difference among the bias estimates gained with RSS, RAOBCORE, and RICH is 0.2 K in Fig. 9b. The radiosonde data adjusted with RICH show similar trends as RAOBCORE, although the anomaly time series differ by more than 0.1 K. Interestingly, the correspondence between RICH and RSS TS time series is better in Fig. 9b than between RAOBCORE and RSS TS. Despite these uncertainties, we see from Fig. 9b that the unadjusted radiosonde temperatures in the TS layer in the late 1980s are at least 0.25 K too warm in the tropics.

RSS does not provide a TS product before 1986 because of large drifts and data gaps in the early 1980s, but it is possible to construct a TS-equivalent time series from both radiosondes and BG for this period. This allows studying the effect of these difficulties on the BG in 1986.

For this purpose, we have plotted the tropical temperature anomaly differences gained with two versions of RAOBCORE. In version 1.4 (Fig. 7) used throughout this paper, the global mean BG time series has been adjusted as described in Haimberger (2007a) before 1987. In the other version (version 1.3; Fig. 10), no modification of the mean BG has been applied at all. In this figure, one sees a large spurious wiggle in the TS-equivalent BG–RAOBCORE anomaly difference in 1986 and a suspicious cool anomaly of the BG relative to radiosondes and satellite data between 1980 and 1984. These breaks are likely attributable to problems in the assimilation of MSU3 and Stratospheric Sounding Unit (SSU) radiances in ERA-40 (Santer et al. 2004; Uppala et al. 2006). Little overlap and large gradual changes in MSU radiances have led to inaccurate bias estimates from static satellite bias–correction schemes (Harris and Kelly 2001) in ERA-40. Results with adaptive bias-correction schemes (Dee 2005; Auligné and McNally 2006) give hope for more accurate bias estimates under such circumstances.

The spurious behavior of the uncorrected BG in 1986 can trigger some spurious break detections unless the BG is adjusted before application of RAOBCORE. This can be seen from the dip in the unadjusted minus adjusted radiosonde time series (blue curve in Fig. 10) around 1985–86, which appears unrealistic. Therefore, the BG had to be modified as outlined in Haimberger (2007a) before 1987. After 1986, the agreement between the uncorrected BG and the other adjusted datasets is good, and certainly much better than the agreement with unadjusted radiosondes.

Figure 7 shows that the wiggle in the BG in 1986 has been efficiently damped in version 1.4. Also, the dark blue composite adjustment time series is now more realistic, with a decrease around 1986. A comparison of Figs. 7, 10 shows that the trend differences between RAOBCORE versions 1.3 and 1.4 are approximately 0.05 K (decade)−1, which is much smaller than the trend differences between unadjusted and adjusted radiosondes [0.4 K (decade)−1 in the LS; 0.2 K (decade)−1 in the TS]. The differences between RICH and RAOBCORE, version 1.4, in Fig. 7 are very small as well. Therefore, the overall cooling of the unadjusted radiosonde temperatures in the TS layer in the 1980s is very likely attributable mainly to large breaks in the radiosonde time series. This conclusion is also consistent with Fig. 9.

In RAOBCORE version 1.2, dataset documented in Haimberger (2007a), the global mean BG has been drawn toward the unadjusted radiosonde dataset also after 1986. The global mean time series from unadjusted radiosonde data have been found to be rather inhomogeneous during the late 1980s and early 1990s by many authors (e.g., Sherwood et al. 2005; Randel and Wu 2006), which implies that spurious inhomogeneities have been added to the BG by this procedure. Since then, Haimberger (2006) has shown that such a BG modification reduces the good consistency of the global and tropical mean BG with the UAH and RSS satellite products after 1987. This alone would not have been sufficient to abandon RAOBCORE version 1.2. The present paper shows that it would also reduce the consistency of the BG with the RICH dataset, which is built of homogeneous pieces of radiosonde time series. Therefore, the global belt–mean trend estimates given in Haimberger (2007a) for version 1.2 are no longer considered valid by the authors. This does not imply that we consider the BG after 1986 as perfect. Because no inhomogeneities could be attributed to it, it has been decided to leave it unadjusted in the present paper.

6. Zonal belt–mean trends and trend profiles

Table 1 summarizes trends from LS- and TS-equivalent layers. In general, the agreement between the two new radiosonde datasets and satellite data is better than for any other homogenized radiosonde dataset published so far (see also Karl et al. 2006; Arguez 2007). Tropical mean trends from radiosonde data adjusted with RICH, which have been gained with break estimates that are strictly independent of satellite data, are practically equal to trends from the UAH dataset in the LS layer. In the global mean, RICH trends show 0.1 K (decade)−1 more cooling because there are still relatively large trend discrepancies in the polar regions, especially Antarctica. The uncertainties involved are discussed below.

The vertical trend profiles from RAOBCORE- and RICH-adjusted radiosonde time series for the global mean and the tropical belt mean are shown in Fig. 11. The tropical trends in Fig. 11a show a maximum around 300 hPa. The maximum warming trend of 0.35 K (decade)−1 in 200–300 hPa in the tropics seems quite high compared to the estimates of surface temperature trends from the HadCRUT version 3.0 dataset [0.15 K (decade)−1, with uncertainties of a few hundredth K (decade)−1; Brohan et al. 2006]. However, such a value is broadly consistent with thermodynamic arguments put forward by Santer et al. (2005) and Held and Soden (2006). These authors predict that surface temperature trends in the tropics are amplified by factors of 1.5–2.5 in the upper troposphere. The trend profile from RICH seems even more consistent with the findings of Santer et al. (2005). It shows a broad maximum with a peak of 0.2 K (decade)−1 in 250 hPa. This maximum is gained from trends of more than 70 stations in the tropics and is quite robust. It is more than supported by the TS-layer mean trends from satellites in Table 1, which shows more warming than RAOBCORE and RICH from 1987 onward. The RSS TS trends would thus be consistent with an even stronger warming maximum in the tropical upper troposphere. Given these results, it appears very likely that the tropical upper-tropospheric trends from radiosondes are stronger than the surface trends, although the range of values [0.2–0.35 K (decade)−1] is still large. Figure 11 also shows the effect of using more (30) neighboring stations in RICH. The upper-tropospheric trend maximum is even larger in this case, but the stratospheric cooling becomes also stronger. At the 100-hPa level, which is very close to the tropical tropopause, the difference between the two RICH versions is largest. We suspect that stations in the extratropics, where the 100-hPa level is clearly in the stratosphere, are more often included in the larger composites. Their 100-hPa temperatures are in a stratospheric cooling regime, while the tropical stations are still in the upper-tropospheric warming regime. This inevitably leads to biased estimates of shifts when using composites.

The trend minimum in 500–700 hPa in the RAOBCORE-adjusted radiosonde data needs further study, although it occurs also in other radiosonde datasets (Thorne et al. 2005) and is also consistent with MSU lower tropospheric (LT) trend estimates of Christy et al. (2007). It is much weaker or absent in the RICH estimates and disappears if the trend period is restricted to 1987–2006. There are indications that the minimum is related to the change from Philips RS4 MKIII to Vaisala RS80 radiosondes, which was performed in Australia, Malaysia, and several surrounding islands in the late 1980s. Adjustment to this change is challenging because of the lack of neighboring reference stations and also because of the strong El Niño–La Niña events occurring in 1987–88 that caused large temperature anomalies in this region. A more thorough analysis, including a comparison with surface temperature, MSU midtropospheric (MT), and MSU LT products especially in this region is needed before the minimum can be interpreted.

The HadAT2 trend profile is found between the unadjusted radiosondes and the RICH estimate. The profile also shows a weak warming maximum in the upper troposphere. McCarthy et al. (2008) and Titchner et al. (2008) have shown with ensemble experiments that an automatic neighbor composite homogenization method similar to HadAT2 tends to remove only about half of the artificial bias, particularly in cases where the bias is pervasive and insufficient breakpoint metadata are available. It appears therefore highly likely that the HadAT2 trend profile underestimates the upper-tropospheric warming.

There is an indication of an upper-tropospheric warming maximum even in the global mean profiles of Fig. 11b. It is more evident in the RAOBCORE trend profile. In the RICH profile, it may be masked by the likely too-strong cooling in the southern extratropics (see Figs. 5, 8d).

a. Uncertainty of RICH and RAOBCORE estimates

Assessing the uncertainty of upper-atmospheric data and statistics derived from these data is a challenging task (Thorne et al. 2005; Karl et al. 2006). Conventional internal error estimates derived from radiosonde and satellite data samples only capture the sampling uncertainty. They yield too-small values because the biases due to the measurement technology of radiosonde data and the processing of the satellite data are large and change over time. This so-called structural uncertainty (Thorne et al. 2005) is much harder to assess because it requires the realization of different plausible processing procedures for the available raw data. This can be implemented either by comparison of datasets from different groups and instrument platforms or by variation of known uncertain parameters of a given method (“exploration of the parameter space,” as exemplified, e.g., by McCarthy et al. 2008). In this paper, we have shown results from two realizations of RAOBCORE and two realizations of RICH, which are summarized in Fig. 11 and which should give a first impression of the overall uncertainties involved.

From the differences between RICH and RAOBCORE, version 1.4, in the belt mean LS and TS time series in Figs. 5 –9 and taking into account the results of Figs. 10, 11 for RAOBCORE, version 1.3, we conclude that the uncertainties for the trend values given in Table 1 for both the LS and TS layers are not more than ±0.15 K (decade)−1 for both the global and the tropical means. This value is derived from the fact that the LS–TS-equivalent differences among these three datasets are not larger than 0.1 K (decade)−1 in all the plots shown, with the exception of the southern extratropics. The trend-sampling uncertainty in the trend estimates has been estimated with a bootstrap method (Lahiri 2003) and is 0.07 K (decade)−1 for TS and 0.1 K (decade)−1 for LS. Given this uncertainty value of ±0.15 K (decade)−1, we note that BG, UAH, and RSS trend estimates lie within the uncertainty bounds given for RICH and RAOBCORE. The trend differences between RICH and RAOBCORE in Fig. 11 are below 0.1 K (decade)−1 at most pressure levels below 30 hPa as well.

The uncertainty value of ±0.15 K (decade)−1 is only half of that given in Haimberger (2007a). This reduction has been achieved (i) by identification and adjustment of the BG inhomogeneity in 1986, (ii) by demonstrating the good consistency between the BG and the datasets derived from MSU radiances, especially in the late 1980s and early 1990s, and (iii) by the ability to trace the odd behavior of radiosonde data during this period to large breaks in individual radiosonde time series. Results achieved with different break-estimation procedures (RICH) corroborated the LS and TS trend estimates from RAOBCORE and are clearly within this uncertainty range. The trend uncertainty estimate is consistent with an uncertainty of ±0.3 K for the tropical and global mean bias of the radiosonde temperatures in the 1980s in Figs. 6, 7, which is estimated as 1.0 ± 0.3 K (tropical) and 0.6 ± 0.3 K (global) for the LS layer. They are estimated as 0.6 ± 0.3 K (tropical) and 0.4 ± 0.3 K (global) for the TS layer. The global uncertainties are not smaller than the tropical uncertainties because of the discrepancies found in the southern extratropics.

The abovementioned “exploration of the parameter space” of RAOBCORE and RICH is beyond the scope of this paper, but it is in preparation (Sperka et al. 2008, unpublished manuscript). This will also involve a RICH version that uses composites of background departures from neighboring stations instead of composites of anomalies from neighboring stations. It is also planned to assess the sensitivity of RICH to the breakpoint dataset provided. The number and location of breakpoints detected by RAOBCORE depend on uncertain thresholds that may be set more conservative or more liberal.

7. Conclusions and outlook

This paper has reported about a comprehensive effort to detect and adjust radiosonde temperature biases in the lower-stratospheric and upper-tropospheric layers. The effort is comprehensive in the sense that radiosonde data, reanalysis products, and MSU data from two channels have been intercompared, in that two new radiosonde adjustment methods have been used, and in that the comparison results are available for every single station as well as for global means. A large number of plots can be viewed online (http://www.univie.ac.at/theoret-met/research/RAOBCORE/). The various station operators and other researchers are invited to compare the adjustment results with their knowledge of radiosonde station histories.

We have documented some new evidence of large, systematic, tropical mean radiosonde temperature biases in the mid-1980s on the order of 1 K in lower-stratospheric (MSU LS equivalent) layers. This bias compared to MSU satellite data has been substantially reduced until the early 1990s with better radiation shielding and better temperature sensors. While the reduction of the bias appears gradual in the tropical or global mean, it is mainly caused by large and systematic shifts toward cooler temperatures at individual radiosonde sites, which often coincide with station metadata events.

The alternative explanation for this change—gradual changes of biases in MSU satellite data—can be excluded, since these would be visible as gradual changes in satellite minus radiosonde time series at most stations, which is not the case.

Two new homogenized datasets have been used in this paper. Haimberger (2007a) has introduced an automatic homogenization approach called Radiosonde Observation Correction using Reanalyses (RAOBCORE) and proved its ability to improve spatial trend consistency and to eliminate spurious changes in the day–night temperature differences. This is desirable but tells little about the effect of the adjustments on global trends. The present paper has provided new evidence for the large stratospheric and upper-tropospheric radiosonde temperature bias in the 1980s compared to satellite data and the ability of RAOBCORE to adjust the bias.

Because one may criticize that the RAOBCORE homogenization method is not independent of satellite data, it has been attempted to adjust the radiosonde temperatures also utilizing an alternative method that uses neighboring radiosondes only. This is quite challenging, as may be seen from the work of Sherwood (2007) and McCarthy et al. (2008). They have shown with several ensemble experiments that automated homogenization methods based on neighbor intercomparison often leave substantial residual biases in the adjusted datasets, unless good metadata are available.

This paper has introduced an automatic algorithm called Radiosonde Innovation Composite Homogenization (RICH), which uses neighbor composites of radiosonde temperature anomalies as reference. It is important to stress that break-size estimation in RICH is performed with radiosonde data only and is thus independent of satellite data. However, RICH does not detect breakpoints themselves, but uses the breakpoint dates found with RAOBCORE as metadata. With these metadata, it appears possible to calculate nearly unbiased reference time series from a few neighboring radiosonde stations for each suspect breakpoint. We have demonstrated that RICH removes most of the global mean cooling bias of radiosonde temperatures compared to the satellite data. Figure 11 indicates that the algorithm works well not only for the LS and TS layers but also for tropospheric layers.

While the evidence of a warm bias in radiosonde temperatures in the 1980s provided here is not based on agreed transfer standards (Thorne et al. 2005), it appears clear that most of the upper-tropospheric trend discrepancy between radiosonde data and climate models can be explained by this warm bias. Both RAOBCORE, version 1.4, and RICH data show a robust upper-tropospheric warming maximum in the tropics. Therefore, both datasets support the arguments of Santer et al. (2005) and Thorne et al. (2007) that the apparent inconsistency in the vertical profile of tropical temperature trends between earlier homogenized radiosonde datasets (HadAT2; RATPAC) and satellite temperature products is to a large fraction caused by residual biases in these radiosonde observation time series. The remaining small cooling of the layer mean LS and TS trends from radiosondes relative to RSS satellite products is insignificant (i.e., within the given uncertainty range). While no climate model data are included in the present intercomparison, we note that the temperature trends from RICH–RAOBCORE version 1.4 are more consistent with trends from recent climate model runs than earlier radiosonde datasets. In the tropical upper troposphere, where the predicted amplification of surface trends is largest, there is no significant discrepancy between trends from RICH–RAOBCORE version 1.4 and the range of temperature trends from climate models. This result directly contradicts the conclusions of a recent paper by Douglass et al. (2007).

However, the contribution to the ongoing discussion on tropical upper-air temperatures is only one component of the present paper. The intercomparison also helped in the assessment of the quality of the ERA-40 background forecast (BG) time series. The comparison of satellite-equivalent BG temperatures with satellite data and satellite-equivalent-adjusted radiosonde data has proven efficient in revealing problems with a particular channel, such as TS in 1986. It has, however, also shown good consistency of the BG with these datasets thereafter. The analysis in the present paper shows that there is good reason to leave the BG uncorrected after 1986, when it is used as input for RAOBCORE, but that a modification as applied in Haimberger (2007a) is needed before 1987. This suggestion has been implemented in the present version 1.4 of RAOBCORE. It may be interesting to also use background forecasts from other reanalyses [e.g., the recently completed JRA-25 reanalysis (Onogi et al. 2007)]. This may help to further constrain the uncertainties involved in the use of background forecasts as a reference for homogenization.

To estimate the uncertainty of the homogeneity adjustments applied to radiosonde time series, we have used not only the BG as reference for homogenization of MSU-equivalent layer mean temperatures but also the UAH and RSS MSU datasets as well as radiosonde neighbor composites. This consistency check yielded almost identical homogeneity adjustments for the larger breaks and therefore relatively small uncertainty. Although it now seems possible to use UAH and RSS data for radiosonde temperature homogenization, the BG remains the reference of choice because it has much better temporal (6 h) horizontal and vertical (1.12°; 60+ levels) resolution than the MSU datasets and also goes back to 1958.

The good break-detection efficiency of RAOBCORE shows how rewarding it is to analyze the background departures from climate data–assimilation systems such as ERA-40. We welcome data assimilation efforts from the early 1930s onward, such as performed by Compo et al. (2006). Such pilot reanalyses will facilitate the homogenization of available early radiosonde data (Bronnimann 2003) and should be strongly supported for this and other reasons. Recently, the RAOBCORE method has been successfully applied to radiosonde wind data as well (Gruber and Haimberger 2008).

The new RICH method has performed very well in this intercomparison and its results may prove even more defensible than RAOBCORE results in the long run, because they are independent of satellite data. We expect that it can perform even better, especially in the southern extratropics, if we put more effort into the determination of the breakpoint dates. The RAOBCORE breakpoints do not appear particularly reliable there.

The RAOBCORE and RICH datasets have been designed to be useful as input for future climate data–assimilation efforts. At the time of this writing, RAOBCORE adjustments are assimilated in the ERA–Interim reanalysis (Uppala 2007; Simmons et al. 2007). A comparison between an assimilation run with radiosonde bias correction using RAOBCORE adjustments and a run without bias correction in the year 1989 shows that the mean radiosonde temperature background departures in the upper troposphere and stratosphere up to 10 hPa are substantially smaller if RAOBCORE data are used. The number of rejected observations is reduced as well, especially in the tropics. The impact on the resulting analyzed temperatures is on the order of 0.2 K in the upper troposphere and stratosphere (Haimberger 2007b), which is certainly not negligible when trying to calculate trends from reanalysis data. In earlier years, when data from only one or no polar-orbiting satellite were available, the impact of radiosonde temperature biases was likely larger. In the tropics, a cool bias in the upper-tropospheric radiosonde measurements may cause artificially increased precipitation in reanalyses (Held and Soden 2006) unless it is adjusted. As long as radiosonde temperature biases cannot be estimated reliably within the data-assimilation process itself (Dee 2005), we recommend the available new adjusted radiosonde temperature data be used as input in climate data–assimilation efforts.

Acknowledgments

This work has been funded by project P18120-N10 of the Austrian Fonds zur Förderung der wissenschaftlichen Forschung (FWF), and in its early stages, by the EC Contract MEIF-CT-2003-503976. It has profited from collaboration with ECMWF within the special project “Homogenization of the global radiosonde temperature and wind dataset” and with the reanalysis section (S. Uppala and D. Dee 2008, personal communication) in particular. Comments from P. Thorne, A. Simmons, and three anonymous reviewers led to substantial improvements. Carl Mears from Remote Sensing Systems, Inc., has provided weighting functions for the TS layer. We thank the UAH group and Remote Sensing Systems, Inc., for making their satellite data available, and the Met Office for providing the HadAT2 data.

REFERENCES

  • Allen, R. J., , and S. C. Sherwood, 2007: Utility of radiosonde wind data in representing climatological variations of tropospheric temperature and baroclinicity in the western tropical pacific. J. Climate, 20 , 52295243.

    • Search Google Scholar
    • Export Citation
  • Arguez, A., 2007: Supplement to State of the Climate in 2006. Bull. Amer. Meteor. Soc., 88 , s1s135.

  • Auligné, T., , and A. McNally, 2006: Adaptive bias correction of satellite data at ECMWF. Proc. ECMWF–NWP-SAF Workshop on Bias Estimation and Correction in Data Assimilation, Reading, United Kingdom, ECMWF, 127–142.

    • Search Google Scholar
    • Export Citation
  • Brohan, P., , J. J. Kennedy, , S. F. B. Tett, , and P. D. Jones, 2006: Uncertainty estimates in regional and global observed temperature changes: A new data set from 1850. J. Geophys. Res., 111 .D12106, doi:10.1029/2005JD006548.

    • Search Google Scholar
    • Export Citation
  • Bronnimann, S., 2003: A historical upper air-data set for the 1939–44 period. Int. J. Climatol., 23 , 769791.

  • Cardinali, C., , S. Pezzuli, , and E. Andersson, 2004: Influence matrix diagnostic of a data assimilation system. Quart. J. Roy. Meteor. Soc., 124 , 17831807.

    • Search Google Scholar
    • Export Citation
  • Charlton, A. J., , and L. M. Polvani, 2007: A new look at stratospheric sudden warmings. Part I: Climatology and modeling benchmarks. J. Climate, 20 , 449469.

    • Search Google Scholar
    • Export Citation
  • Charlton, A. J., and Coauthors, 2007: A new look at stratospheric sudden warmings. Part II: Evaluation of numerical model simulations. J. Climate, 20 , 470488.

    • Search Google Scholar
    • Export Citation
  • Christy, J. R., , W. B. Norris, , W. D. Braswell, , and D. E. Parker, 2003: Error estimates of version 5.0 of MSU–AMSU bulk atmospheric temperatures. J. Atmos. Oceanic Technol., 20 , 613629.

    • Search Google Scholar
    • Export Citation
  • Christy, J. R., , W. B. Norris, , R. W. Spencer, , and J. J. Hnilo, 2007: Tropospheric temperature change since 1979 from tropical radiosonde and satellite measurements. J. Geophys. Res., 112 .D06102, doi:10.1029/2005JD006881.

    • Search Google Scholar
    • Export Citation
  • Compo, G. P., , J. S. Whitacker, , and P. D. Sardeshmukh, 2006: Feasibility of a 100-year reanalysis using only surface pressure data. Bull. Amer. Meteor. Soc., 87 , 175190.

    • Search Google Scholar
    • Export Citation
  • Dee, D. P., 2005: Bias and data assimilation. Quart. J. Roy. Meteor. Soc., 131 , 33233343.

  • Douglass, D. H., , J. R. Christy, , B. D. Pearson, , and S. F. Singer, 2007: A comparison of tropical temperature trends with model predictions. Int. J. Climatol., 28 .doi:10.1002/joc.1651.

    • Search Google Scholar
    • Export Citation
  • Durre, I., , R. S. Vose, , and D. B. Wuertz, 2006: Overview of the Integrated Global Radiosonde Archive. J. Climate, 19 , 5368.

  • Free, M., , and D. J. Seidel, 2005: Causes of differing temperature trends in radiosonde upper air data sets. J. Geophys. Res., 110 .D07101, doi:10.1029/2004JD005481.

    • Search Google Scholar
    • Export Citation
  • Gaffen, D. J., 1996: A digitized metadata set of global upper-air station histories. NOAA Tech. Memo. ERL ARL-211, 38 pp.

  • Gruber, C., , and L. Haimberger, 2008: On the homogeneity of radiosonde wind time series. Meteor. Z., in press.

  • Haimberger, L., 2006: Towards temporally homogeneous evaluations of the observed global atmospheric circulation. Habilitationsschrift (habilitation thesis), University of Vienna, 61 pp. [Available online at http://homepage.univie.ac.at/leopold.haimberger/raobcone1-Dateien/Habilschrift_full.pdf.].

  • Haimberger, L., 2007a: Homogenization of radiosonde temperature time series using innovation statistics. J. Climate, 20 , 13771403.

  • Haimberger, L., 2007b: Homogenization of the global radiosonde temperature and wind dataset. ECMWF Special Project Rep. 2007, 8 pp. [Available online at http://www.ecmwf.int/about/special_projects/haimberger_temp_wind_dataset/report_2007.pdf.].

  • Harris, B. A., , and G. A. Kelly, 2001: A satellite radiance bias correction scheme for data assimilation. Quart. J. Roy. Meteor. Soc., 127 , 14531468.

    • Search Google Scholar
    • Export Citation
  • Held, I. M., , and B. J. Soden, 2006: Robust responses of the hydrological cycle to global warming. J. Climate, 19 , 56865699.

  • Karl, T. R., , S. J. Hassol, , C. D. Miller, , and W. L. Murray, and Eds., 2006: Temperature trends in the lower atmosphere: Steps for understanding and reconciling differences. U.S. Climate Change Science Program and the Subcommittee on Global Change Research Final Rep. Synthesis and Assessment Product 1.1, 180 pp. [Available online at http://www.climatescience.gov/Library/sap/sap1-1/finalreport/sap1-1-final-all.pdf.].

  • Lahiri, S. K., 2003: Resampling Methods for Dependent Data. Springer, 374 pp.

  • Lanzante, J. R., , S. A. Klein, , and D. J. Seidel, 2003: Temporal homogenization of monthly radiosonde temperature data. Part II: Trends, sensitivities, and MSU comparison. J. Climate, 16 , 241262.

    • Search Google Scholar
    • Export Citation
  • Limpasuvan, V., , D. W. J. Thompson, , and D. L. Hartmann, 2004: The life cycle of the Northern Hemisphere sudden stratospheric warmings. J. Climate, 17 , 25842596.

    • Search Google Scholar
    • Export Citation
  • McCarthy, M. P., , H. Titchner, , P. Thorne, , L. Haimberger, , and D. E. Parker, 2008: Assessing bias and uncertainty in the HadAT-adjusted radiosonde climate record. J. Climate, 21 , 817832.

    • Search Google Scholar
    • Export Citation
  • Mears, C. A., , M. C. Schabel, , and F. J. Wentz, 2003: A reanalysis of the MSU channel 2 tropospheric temperature record. J. Climate, 16 , 36503664.

    • Search Google Scholar
    • Export Citation
  • Mears, C. A., , C. E. Forest, , R. W. Spencer, , R. S. Vose, , and R. W. Reynolds, 2006: What is our understanding of the contribution made by observational or methodological uncertainties to the previously reported vertical differences in temperature trends? Temperature trends in the lower atmosphere: Steps for understanding and reconciling differences, U.S. Climate Change Science Program and the Subcommittee on Global Change Research Final Rep. Synthesis and Assessment Product 1.1., 71–88.

  • Onogi, K., and Coauthors, 2007: The JRA-25 reanalysis. J. Meteor. Soc. Japan, 85 , 369432. doi:10.2151/jmsj.85.369.

  • Randel, W. J., , and F. Wu, 2006: Biases in stratospheric and tropospheric temperature trends derived from historical radiosonde data. J. Climate, 19 , 20942104.

    • Search Google Scholar
    • Export Citation
  • Santer, B. D., and Coauthors, 2004: Identification of anthropogenic climate change using a second-generation reanalysis. J. Geophys. Res., 109 .D21104, doi:10.1029/2004JD005075.

    • Search Google Scholar
    • Export Citation
  • Santer, B. D., and Coauthors, 2005: Amplification of surface temperature trends and variability in the tropical atmosphere. Science, 309 , 15511556.

    • Search Google Scholar
    • Export Citation
  • Seidel, D. J., , R. J. Ross, , J. K. Angell, , and G. C. Reid, 2001: Climatological characteristics of the tropical tropopause as revealed by radiosondes. J. Geophys. Res., 106 , 78577878.

    • Search Google Scholar
    • Export Citation
  • Sherwood, S. C., 2007: Simultaneous detection of climate change and observing biases in a network with incomplete sampling. J. Climate, 20 , 40474062.

    • Search Google Scholar
    • Export Citation
  • Sherwood, S. C., , J. Lanzante, , and C. Meyer, 2005: Radiosonde daytime biases and late-20th century warming. Science, 309 , 15561559. doi:10.1126/science.1115640309.

    • Search Google Scholar
    • Export Citation
  • Simmons, A., , S. Uppala, , D. Dee, , and S. Kobayashi, 2007: The ERA interim reanalysis. ECMWF Newsletter, No. 110, ECMWF, Reading, United Kingdom, 25–35.

  • Sperka, S., 2007: Homogeneity adjustments of radiosonde temperature time series using composites of innovations from ERA-40. M.S. thesis, Department of Meterology and Geophysics, University of Vienna, 63 pp. [Available online at http://www.univie.ac.at/img-wien/dipldiss/dipl/DA_Sperka.pdf.].

  • Steiner, A. K., , G. Kirchengast, , M. Borsche, , U. Foelsche, , and T. Schoengassner, 2008: A multi-year comparison of lower stratospheric temperatures from CHAMP radio occultation data with MSU/AMSU records. J. Geophys. Res., 112 .D22110, doi:10.1029/2006JD008283.

    • Search Google Scholar
    • Export Citation
  • Thorne, P. W., , D. E. Parker, , S. F. B. Tett, , P. D. Jones, , M. McCarthy, , H. Coleman, , and P. Brohan, 2005a: Revisiting radiosonde upper air temperatures from 1958 to 2002. J. Geophys. Res., 110 .D18105, doi:10.1029/2004JD005753.

    • Search Google Scholar
    • Export Citation
  • Thorne, P. W., , D. E. Parker, , J. R. Christy, , and C. A. Mears, 2005b: Uncertainties in climate trends: Lessons from upper-air temperature records. Bull. Amer. Meteor. Soc., 86 , 14371442.

    • Search Google Scholar
    • Export Citation
  • Thorne, P. W., and Coauthors, 2007: Tropical vertical temperature trends: A real discrepancy? Geophys. Res. Lett., 34 .L16702, doi:10.1029/2007GL029875.

    • Search Google Scholar
    • Export Citation
  • Titchner, H., , P. W. Thorne, , M. McCarthy, , S. F. B. Tett, , L. Haimberger, , and D. E. Parker, 2008: Critically reassessing tropospheric temperature trends from radiosondes using realistic validation experiments. J. Climate, in press.

    • Search Google Scholar
    • Export Citation
  • Trenberth, K. E., , and L. Smith, 2006: The vertical structure of temperature in the tropics: Different flavors of El Niño. J. Climate, 19 , 49564970.

    • Search Google Scholar
    • Export Citation
  • Trenberth, K. E., and Coauthors, 2007: Observations: Surface and atmospheric climate change. Climate Change 2007: The Physical Science Basis, S. Solomon et al., Eds., Cambridge University Press, 235–336.

    • Search Google Scholar
    • Export Citation
  • Untch, A., , M. Miller, , M. Hortal, , R. Buizza, , and P. A. E. M. Janssen, 2006: Towards a global mesoscale model: The high-resolution system T799L91 and T399L62 EPS. ECMWF Newsletter, No. 108, ECMWF, Reading, United Kingdom, 6–13.

  • Uppala, S. M., 2007: From ERA-15 to ERA-40 and ERA-Interim. Proc. ECMWFGEO Workshop on Atmospheric Reanalysis, Reading, United Kingdom, ECMWF, 17–22.

    • Search Google Scholar
    • Export Citation
  • Uppala, S. M., and Coauthors, 2005: The ERA-40 Re-analysis. Quart. J. Roy. Meteor. Soc., 131 , 29613012.

  • Uppala, S. M., , G. Kelly, , B. K. Park, , P. Kallberg, , and A. Untch, 2006: Experience in estimation of biases in ECMWF reanalyses. Proc. ECMWFNWP–SAF Workshop on Bias Estimation and Correction in Data Assimilation, Reading, United Kingdom, ECMWF, 207–224.

    • Search Google Scholar
    • Export Citation

Fig. 1.
Fig. 1.

Decadal temperature trends at radiosonde stations for period 1979–2006, derived (top left) from raw data, (top right) from data homogenized with RAOBCORE, (bottom left) from RSS MSU data and (bottom right) from RICH. Trends are 0000 + 1200 UTC averages, if both time series are available; otherwise, they are derived from either 0000 or 1200 UTC ascents. Here, 400 stations have complete enough time series (at least 26 out of 28 yr of data up to the 30-hPa level). “Cost” quantifies spatial heterogeneity—the smaller the better. Acronyms are explained in section 2.

Citation: Journal of Climate 21, 18; 10.1175/2008JCLI1929.1

Fig. 2.
Fig. 2.

Time series of monthly mean 0000 UTC temperature anomaly differences with respect to RSS MSU satellite data (a) for the LS and (b) for the TS layer at station Yap (91413, tropical West Pacific). Break dates detected by RAOBCORE are 19820304, 19870904, 19900720, 19951201, and 20021114. Differences are RAOBCORE − RSS (light blue), UAH − RSS (orange), BG − RSS (green), RICH − RSS (pink), HadAT2 − RSS (dark red), and unadjusted (RS) data − RSS (dark blue). Triangles at bottom indicate documented radiosonde instrumentation changes. Corresponding decadal trend differences 1979–2006 (LS) and 1987–2006 (TS) are plotted on the rhs. No TS data are available before 1987.

Citation: Journal of Climate 21, 18; 10.1175/2008JCLI1929.1

Fig. 3.
Fig. 3.

Same as Fig. 2, but for radiosonde station Bethel (70219, Alaska). Break dates detected by RAOBCORE are 19840404, 19890701, 19920222, 19950724, 19970725, and 19990814: (a), (b) 0000 UTC, (c), (d) 1200 UTC.

Citation: Journal of Climate 21, 18; 10.1175/2008JCLI1929.1

Fig. 4.
Fig. 4.

Same as Fig. 2, but for radiosonde station Jokioinen (2963, Finland) at 0000 UTC. Break dates detected by RAOBCORE are 19811124, 19860215, and 20050303. Largest differences occur between 1989 and 1993 in January (i.e., in the deepest polar winter).

Citation: Journal of Climate 21, 18; 10.1175/2008JCLI1929.1

Fig. 5.
Fig. 5.

Zonal mean decadal temperature trends (a) for the LS layer during 1979–2006 and (b) for the TS layer during 1987–2006. Zonal means have been calculated from 10° × 10° box means where at least two radiosonde stations were available. Diamonds are plotted at the center of 10° belts, e.g., 5° for 0°–10°N. Insufficient data available for 90°–70°S and 80°–90°N.

Citation: Journal of Climate 21, 18; 10.1175/2008JCLI1929.1

Fig. 6.
Fig. 6.

Time series of global mean MSU LS–TS temperature anomaly differences RSS − RAOBCORE (red), UAH − RAOBCORE (orange), ERA-40 BG − RAOBCORE (green), RICH − RAOBCORE (pink), HadAT2 − RAOBCORE (dark red), and unadjusted radiosonde − RAOBCORE (dark blue). Corresponding trend differences are shown on the rhs. All global means are from data subsampled at radiosonde stations. Note that in contrast to the station intercomparisons, the anomaly differences are now with respect to RAOBCORE. RAOBCORE trends for the LS and TS layers are −0.4 K (decade)−1 and −0.04 K (decade)−1, respectively. The large variance of the HadAT2 − RAOBCORE difference series is caused by the much coarser spatial sampling of the HadAT2 dataset south of the equator.

Citation: Journal of Climate 21, 18; 10.1175/2008JCLI1929.1

Fig. 7.
Fig. 7.

Same as Fig. 6, but for the tropics (20°S–20°N). The figure has been derived from 42 stations in the tropics with 26 out of 28 yr of data since 1979. Difference of dark blue curve in the 1980s minus the 2000s yields RAOBCORE bias estimates (approx 1 K in the LS; approx 0.6 K in the TS). RAOBCORE trends for the LS and TS layers are −0.34 K (decade)−1 and −0.02 K (decade)−1, respectively.

Citation: Journal of Climate 21, 18; 10.1175/2008JCLI1929.1

Fig. 8.
Fig. 8.

Decadal trend differences with respect to RAOBCORE for the (top) LS layer and (bottom) TS layer. Differences are calculated (a) for the tropics during nighttime [0000 UTC between 90°W and 90°E, 1200 UTC between 90°E and 90°W, RAOBCORE trends −0.32 (−0.06) K (decade)−1 for the LS (TS) layer], (b) for the tropics during daytime [RAOBCORE trends −0.36 (−0.02) K (decade)−1 for the LS (TS) layer], (c) for the northern extratropics (20°–90°N) daily means [RAOBCORE trends −0.37 (0.07) K for the LS (TS) layer], (d) for the southern extratropics (20°–90°S) daily means [RAOBCORE trends −0.48 (−0.13) K (decade)−1 for the LS (TS) layer].

Citation: Journal of Climate 21, 18; 10.1175/2008JCLI1929.1

Fig. 9.
Fig. 9.

Mean effect of radiosonde temperature adjustments in the tropics (20°N–20°S) for the (a) LS and (b) TS layer. The figure in the title is the number of radiosonde stations that have been adjusted. Thick curves show effect of large breaks (>0.5 K for the LS layer; >0.3 K for the TS layer), and thin curves show effect of breaks smaller than these thresholds. Blue curves show effect of adjustments estimated by RAOBCORE. Red curves are effect of adjustments calculated from RSS-unadjusted radiosonde differences using the same time intervals as RAOBCORE. Practically the same curve as with RSS LS data can be obtained with UAH LS data (not shown). Pink curves are calculated from adjustments estimated with RICH.

Citation: Journal of Climate 21, 18; 10.1175/2008JCLI1929.1

Fig. 10.
Fig. 10.

Same as Fig. 7, but with anomaly differences with respect to RAOBCORE, version 1.3, where the BG has not been corrected before radiosonde adjustment at all, not even before 1987. HadAT2 series is omitted here for clarity and because it is already plotted in Fig. 7.

Citation: Journal of Climate 21, 18; 10.1175/2008JCLI1929.1

Fig. 11.
Fig. 11.

Vertical temperature trend profiles (a) for the tropics (20°S–20°N) and (b) for the global mean. Thick solid curve is standard RICH estimate using standard eight reference stations. Thick dashed–double dotted curve is RICH estimate using 30 reference stations. Thin solid curve is RAOBCORE, version 1.4, estimate, thin dashed is RAOBCORE, version 1.3, and dotted is from unadjusted radiosonde data. HadAT2 profiles (thin dashed–dotted) are estimated from less available radiosondes and are included for reference. Corresponding surface temperature trends from HadCRUT, version 3.0 are denoted with x symbols.

Citation: Journal of Climate 21, 18; 10.1175/2008JCLI1929.1

Table 1.

Decadal linear least squares trends in K (decade)−1 for latitudinal belts and the globe for the MSU LS and TS layer. RSS is RSS sampled at radiosonde stations, RSS ALL is gridded original RSS data, UAH is UAH sampled at radiosonde stations, UAH all is gridded original UAH data, BG is ERA-40 + ECMWF operational background, with global mean adjustment applied before 1987, UNADJ is raw radiosonde data, ADJ V1.4 and ADJ V1.3 are from data adjusted with RAOBCORE versions 1.4 and 1.3, respectively, RICH obs is RICH-adjusted data, HadAT2 is HadAT2-adjusted data. Sampling uncertainty estimated from belt mean time series with the bootstrap method is ±0.1 K (decade)−1 for LS and ±0.07 K (decade)−1 for TS. HadAT2 data have not been sampled in the same way as in Arguez (2007), where the tropical HadAT2 estimate is −0.55 K (10 yr)−1.

Table 1.
Save