Sensitivity of tropospheric and stratospheric temperature trends to radiosonde data quality

Radiosonde data have been used, and will likely continue to be used, for th e detection of temporal trends in tropospheric and lower-stratospheric temp erature. However, the data are primarily operational observations, and it i s not clear that they are of sufficient quality for precise monitoring of c limate change. This paper explores the sensitivity of upper-air temperature trend estimates to several data quality issues. Many radiosonde stations do not have even moderately complete records of mo nthly mean data for the period 1959-95. In a network of 180 stations (the c ombined Global Climate Observing System Baseline Upper-Air Network and the network developed by J. K. Angell), only 74 stations meet the data availabi lity requirement of at least 85% of nonmissing months of data for troposphe ric levels (850-100 hPa). Extending into the lower stratosphere (up to 30 h Pa), only 22, stations have data records meeting this requirement for the s ame period, and the 30-hPa monthly data are generally based on fewer daily observations than at 50 hPa and below. These networks show evidence of stat istically significant tropospheric warming, particularly in the Tropics, an d stratospheric cooling for the period 1959-95. However, the selection of d ifferent station networks can cause network-mean trend values to differ by up to 0.1 K decade(-1). The choice of radiosonde dataset used to estimate trends influences the res ults. Trends at individual stations and pressure levels differ in two indep endently produced monthly mean temperature datasets. The differences are ge nerally less than 0.1 K decade(-1), but in a few cases they are larger and statistically significant at the 99% confidence level. These cases are due to periods of record when one dataset has a distinct bias with respect to t he other. The statistical method used to estimate linear trends has a small influence on the result. The nonparametric median of pairwise slopes method and the parametric least squares linear regression method tend to yield very simila r, but not identical, results with differences generally less than +/-0.03 K decade(-1) for the period 1959-95. However, in a few instances the differ ences in stratospheric trends for the period 1970-95 exceed 0.1 K decade(-1 ). Instrument changes can lead to abrupt changes in the mean, or change-points , in radiosonde temperature data records, which influence trend estimates. Two approaches to removing change-points by adjusting radiosonde temperatur e data were attempted. One involves purely statistical examination of time series to objectively identify and remove multiple change-points. Methods o f this type tend to yield similar results about the existence and timing of the largest change-points, but the magnitude of detected change-points is very sensitive to the particular scheme employed and its implementation. Th e overwhelming effect of adjusting time series using the purely statistical schemes is to remove the trends, probably because some of the detected cha nge-points are not spurious signals but represent real atmospheric change. The second approach incorporates station history information to test specif ic dates of instrument changes as potential change-points, and to adjust ti me series only if there is agreement in the test results for multiple stati ons. This approach involved significantly fewer adjustments to the time ser ies, and their effect was to reduce tropospheric warming trends (or enhance tropospheric cooling) during 1959-95 and (in the case of one type of instr ument change) enhance stratospheric cooling during 1970-95. The trends base d on the adjusted data were often statistically significantly different fro m the original trends at the 99% confidence level. The intent here was not to correct or improve the existing time series, but to determine the sensit ivity of trend estimates to the adjustments. Adjustment for change-points c an yield very different time series depending on the scheme used and the ma nner in which it is implemented, and trend estimates are extremely sensitiv e to the adjustments. Overall, trends are more sensitive to the treatment o f potential change-points than to any of the other radiosonde data quality issues explored.