A New Method to Homogenize Atmospheric Radiosonde Daily Temperature Data

Atmospheric temperature (T) is important for studying atmospheric dynamics and quantifying atmospheric warming. Since the 1950s, radiosonde measurements provide a long-term in-situ record of atmospheric temperature, humidity and winds that have been used to constrain historical atmospheric reanalyses. Their applications are, however, hampered by numerous discontinuities in these data arising from changes in instruments, observational practices, and bias correction methods. Previous homogenization efforts have mainly focused on the removal of spurious shifts in the mean without adjustments for spurious changes in the variance, which are evident in many radiosonde time series that could greatly affect the detection of extreme events. Here we attempt to homogenize the twice daily radiosonde T data collected from the IGRA2 (91%) and the ERA assimilation series (9%) since the 1950s over the globe. After many tests, we apply the following steps: 1) constructing T difference series by subtracting monthly T from NOAA 20CR reanalysis and high-frequency T variations from JRA55 reanalysis from the radiosonde T series to remove natural variations; 2) merging the changepoints detected by applying a modified Kolmogorov-Smirnov (K-S) test for shifts in the variation (after removing the mean) and a penalized maximal F test (PMFtest) for shifts in the mean of the difference series, and the merged changepoints were verified using available metadata and by the changepoints detected using the daytime-nighttime T difference series; and 3) removing the discontinuities in the original T series at the detected changepoints using a quantile-matching (QM) algorithm, which adjusts not only the mean but also the variance. Tests showed that our changepoint detection method works well in detecting the shifts in the mean and variation embedded in synthetic time series. Approximately 83% of the detected changepoints match the change dates recorded in metadata and the changepoints display a country-wide pattern to some extent. Our preliminary analysis indicates that the radiosonde T data since the 1950s have approximately 6.8 changepoints on average globally, of which approximately 47% (3.2 changepoints) are due to the spurious shifts in the variance only. These preliminary results imply that the shifts in the variance are non-negligible, and that our new homogenization method is very promising for homogenizing the historical radiosonde temperature data, which are needed for estimating atmospheric humidity variables and quantifying tropospheric warming trends and changes in the lapse rate, as well as for use in the next generation of atmospheric reanalyses.