The tropical tropopause layer in reanalysis data sets 2 3

4European Centre for Medium-Range Weather Forecasts, Reading, RG2 9AX, UK 14 5Japan Meteorological Agency, Tokyo, 100-8122, Japan 15 6 Section for Meteorology and Oceanography, Department of Geosciences, University of Oslo, 16 0315 Oslo, Norway 17 7Laboratoire de Météorologie Dynamique, CNRS/PSL-ENS, Sorbonne University Ecole 18 Polytechnique, France 19 8Climate Prediction Center, National Centers for Environmental Prediction, National Oceanic 20 and Atmospheric Administration, College Park, MD 20740, USA 21

Interannual variability in reanalysis temperatures is best constrained in the upper TTL, with 23 larger differences at levels below the cold point. The reanalyses reproduce the temperature 24 responses to major dynamical and radiative signals such as volcanic eruptions and the QBO. 25 Long-term reanalysis trends in temperature in the upper TTL show good agreement with trends 26 derived from adjusted radiosonde data sets indicating significant stratospheric cooling of 27 around -0.5 to -1 K/decade. At 100 hPa and the cold point, most of the reanalyses suggest small 28 but significant cooling trends of -0.3 to -0.6 K/decade that are statistically consistent with 29 trends based on the adjusted radiosonde data sets. 30 Advances of the reanalysis and observational systems over the last decades have led to a clear 31 improvement of the TTL reanalyses products over time. Biases of the temperature profiles and 32 differences in interannual variability clearly decreased in 2006, when densely sampled radio 33 occultation data started being assimilated by the reanalyses. While there is an overall good 34 agreement, different reanalyses offer different advantages in the TTL such as realistic profile 35 and cold point temperature, continuous time series or a realistic representation of signals of 36 interannual variability. Their use in model simulations and in comparisons with climate model 37 output should be tailored to their specific strengths and weaknesses. 38 39 40 41 42 43 In the tropics, two definitions of the tropopause are widely used: one based on the cold point 23 and one based on the characteristics of the lapse rate. The cold point tropopause is defined as 24 the level at which the vertical temperature profile reaches its minimum (Highwood and Hoskins,25 1998) and air parcels en route from the troposphere to the stratosphere encounter the lowest 26 temperatures. Final dehydration typically occurs at these lowest temperatures, so that the cold 27 point tropopause effectively controls the overall water vapour content of the lower stratosphere 28 (Randel et al., 2004a) and explains its variability (Fueglistaler et al., 2009). While the cold point 29 tropopause is an important boundary in the tropics where upwelling predominates, this 30 definition of the tropopause is irrelevant for water vapor transport into the stratosphere at higher 31 latitudes. The lapse rate tropopause, on the other hand, offers a globally-applicable definition 32 of the tropopause, marking a vertical discontinuity in the static stability. The lapse rate 33 tropopause is defined as the lowest level at which the lapse rate decreases to 2 K km -1 or less, 34 provided that the average lapse rate between this level and all higher levels within 2 km does 35 not exceed 2 K km -1 (World Meteorological Organization, 1957). The tropical lapse rate 36 tropopause is typically ~0.5 km (~10 hPa) lower and ~1 K warmer than the cold point 37 tropopause (Seidel et al., 2001). 38 39 Over recent decades, the thermal characteristics of the TTL and tropopause have been obtained 40 from tropical radiosonde and Global Navigation Satellite System -Radio Occultation (GNSS-41 RO) upper air measurements. Radiosonde profiles offer temperature, wind and air pressure data 42 at a high vertical resolution. However, climate records based on radiosonde data often suffer 43 from spatial inhomogeneities or time-varying biases due to changes in instruments and  Meteorological reanalysis data sets are widely used in scientific studies of atmospheric 6 processes and variability, either as initial conditions for historical model runs or in comparisons 7 with climate model output. Often, they are utilized as "stand-ins" for observations, when the 8 available measurements lack the spatial or temporal coverage needed. Each atmospheric 9 reanalysis system consists of a fixed global forecast model and assimilation scheme. The system 10 combines short-range forecasts of the atmospheric state with available observations to produce 11 best-guess, consistent estimates of atmospheric variables such as temperatures and winds.

12
Spurious changes in the reanalysis fields can arise from changes in the quality and quantity of 13 the observations used as input data, which complicates the analysis of variability and trends.
14 Further discontinuities in reanalysis-based time series can originate from the joining together 15 of distinct execution streams (Fujiwara et al., 2017). 16 17 Among the various TTL characteristics such as composition, radiation budgets and cloud 18 properties, the vertical temperature structure and the position and temperature of the cold point 19 are of particular importance for transport and composition studies. Many off-line chemistry-20 transport models or Lagrangian particle dispersion models are driven by reanalysis data sets 21 (e.g., Chipperfield, 1999;Krüger et al., 2009;Schoeberl et al, 2012). Their representation of the 22 cold point determines how realistically such models simulate dehydration and stratospheric 23 entrainment processes. Process studies of TTL dynamics such as equatorial wave variability are 24 also often based on the TTL temperature structure in reanalysis data sets (e.g., Fujiwara et al.,  A comparison of the reanalysis products available at the end of the 1990s (including ERA-15, 32 ERA-40 and NCEP-NCAR R1) with other climatological datasets showed notable differences 33 in temperatures near the tropical tropopause (Randel et al., 2004b). While the ECWMF 34 reanalyses agreed relatively well with radiosonde observations at 100 hPa, NCEP-NCAR R1 35 showed a warm bias of up to 3 K, probably resulting from low vertical resolution and the use 36 of poorly-resolved satellite temperature retrievals (Fujiwara et al., 2017). Comparisons of 37 winter temperatures at 100 hPa between more recent reanalyses, such as MERRA, NCEP CFSR 38 and ERA-Interim, and Singapore radiosonde observations show better agreement, with 39 reanalyses generally 1-2 K too cold at this level (Schoeberl et al, 2012). While many studies 40 have highlighted the characteristics of individual reanalysis data sets, a comprehensive 41 intercomparison of the TTL among all major atmospheric reanalyses is currently missing.

43
Here, we investigate whether reanalysis data sets reproduce key characteristics of the 44 temperature and tropopauses in the TTL. This work has been conducted as part of the SPARC Project (S-RIP) (Fujiwara et al., 2017) and presents some of the key findings from the S-RIP 2 report Chapter 8 on the TTL. Climatologies of the tropical cold point and lapse rate tropopauses 3 as derived from modern reanalysis data sets are compared to high-resolution radio occultation 4 data (Section 3). We also investigate temporal variability and long-term changes in tropopause 5 levels and temperature within the TTL (Section 4). The observational and reanalysis data sets 6 used in the evaluation are introduced in Section 2, and a discussion and summary of the results 7 are provided in Section 5. 2 Data and methods 10 11 2.1 Observational data sets 12 13 High-resolution observations of the TTL are available from tropical radiosonde stations. 14 However, climate records of radiosonde temperature, height and pressure data often suffer from 15 inhomogeneities or time-varying biases due to changes in instruments or measurement practices 16 (Seidel and Randel, 2006 (Haimberger, 2007) and HadAT (Thorne et al., 2005) as well as the unadjusted, quality-   profiles with the barometric formula, taking into account the lapse rate between levels. For each 37 profile, the cold point and lapse rate tropopause characteristics were identified based on the 38 cold point and WMO criteria, respectively. 39 We also use a daily data set of cold point temperatures obtained from all GNSS-RO missions, 40 gridded on a 5°x5° grid between 30°N and 30°S. For each 5° wide latitude band, we apply a wider range of equivalent depths, since it has been shown that Kelvin waves tend to propagate 1 faster around the tropical tropopause than they do in the troposphere (Kim and Son, 2012). The 2 filtered anomalies represent cold point temperature variability that propagates in the same 3 wavenumber-frequency domain as Kelvin waves, i.e. when the temperature is modulated by 4 Kelvin waves present around the tropopause. The spatial variance of the filtered signals is used 5 to calculate a monthly index as a measure of the amount of Kelvin wave activity in the TTL. 6 The index is calculated as the 1s standard deviation over the filtered anomalies at all spatial 7 grid points. Time periods of enhanced Kelvin wave activity are defined as the months when the 8 index is larger than the long-term mean plus the 1s standard deviation of the whole time series. 9 Based on this definition, we determined 20% of all months to be characterized by enhanced 10 Kelvin wave activity.  13 14 We evaluate eight "full-input" reanalyses, defined as systems that assimilate surface and upper-  . 25 Global temperature fields in the reanalysis data sets are produced by assimilating conventional 26 (surface and balloon), aircraft, and satellite observations. The most important sources of 27 assimilated data for stratospheric temperatures are the microwave and infrared satellite 28 sounders of the TOVS suite  and the ATOVS suite (1998-present). All of the 29 above reanalysis systems assimilate microwave and infrared radiances from these instruments, 30 except for NCEP-NCAR R1 which assimilates temperature retrievals instead. Measurements    Among the observational data sets, radiosonde and GNSS-RO data are our best source of 25 information about the TTL. While the reanalyses assimilate versions of these data, it is not 26 automatic that they reproduce the data exactly. For instance, discrepancies exist between 27 reanalysis stratospheric temperatures and those derived from their radiance input data (Long et compared to radiosonde and GNSS-RO data also leads to differences, especially for derived 2 quantities such as the tropopause location and temperature, which will be investigated in the 3 following evaluations. 4 The reanalysis models resolve the TTL with different vertical resolutions, as illustrated in and 70 hPa ( Fig. 1). We show that reanalysis-based estimates of tropopause temperature, 10 pressure and height compare much better to observations when they are derived from model- 11 level data than when they are derived from pressure-level data (Section 3.1). Another sensitivity 12 study demonstrates that tropopause temperatures directly calculated from monthly-mean fields 13 have a warm bias of 0.5 K compared to tropopause temperatures based on 6-hourly data (not 14 shown here). Therefore, we derive the cold point and lapse rate tropopause characteristics for 15 each reanalysis using model-level data at each grid point at 6-hourly temporal resolution. Zonal

25
Given the strong gradients of temperature and static stability in the TTL, the vertical resolution 26 of the reanalysis data sets is an important factor in cold point and lapse rate tropopause 27 calculations. For each reanalysis, tropopause heights and temperatures can be derived either 28 from model-or pressure-level data (Fig 1). A comparison of the CFSR cold point tropopause 29 based on model-and pressure-level temperature data is shown here to demonstrate the clear 30 advantage of the finer model-level resolution (Fig. 2). The cold point tropopause from CFSR  (1) 15 Here QBO1( ) and QBO2( ) are orthogonal time series representing QBO variations constructed 16 as the first two EOFs of the Freie Universität Berlin (FUB) radiosonde stratospheric winds 17 (Naujokat, 1986). The long-term trends of the reanalyses temperature time series have been derived as the 26 regression coefficient of a linear function that provides the best fit in a least-squares sense. The 27 trend error bars are as the standard error of the slope with an effective sample size. Significance 28 is tested based on two-tailed test with a 95% confidence interval. 3 Temperature and tropopause characteristics 1 2 Tropical mean temperatures from reanalyses at two standard pressure levels (100 hPa and 70 3 hPa) and at the two tropopause levels are compared to radio occultation data for the time period 2002-2010 (Fig. 3). At 100 hPa, reanalysis temperatures agree well with radio occultation data 5 with differences between -0.35 K (too cold; ERA-Interim and ERA5) and 0.43 K (too warm; 6 CFSR). At 70 hPa, the agreement is even better, with differences ranging from -0.29 K (JRA-  Temperature profile comparisons between 140 and 70 hPa at the native model level resolution 22 have been conducted for the five most recent reanalyses (ERA5, ERA-Interim, JRA-55, 23 MERRA-2, CFSR). All reanalyses tend to be colder than the observations in the tropical mean 24 (Fig. 4), but differences are relatively small and the agreement is good overall. CFSR and ERA5 25 agree best with the radio occultation data with mean biases of around -0.06 K and -0.28 K, 26 respectively, averaged over the whole vertical range. ERA-Interim and MERRA-2 agree very 27 well at upper levels but show large deviations near 100 hPa (ERA-Interim; -0.82 K) and below 28 110 hPa (MERRA-2; -0.67 K), respectively. The evaluation demonstrates that temperature 29 comparisons at standard pressure levels (Fig. 3) can be biased by up to 0.5 K, with CFSR Comparing the temperature profiles to the tropopause values ( Fig. 3 and 4) reveals that despite   comparisons to radio occultation data (Fig. 6). All reanalysis data sets produce tropopause 13 levels that are too low and too warm, with the latter related to vertical resolution as explained  based on R1 is both higher and warmer than observed. The best agreement with respect to cold 13 point temperatures is found for ERA5 and ERA-Interim, which are around 0.2 K and 0.4 K 14 warmer than the radio occultation data, respectively. All other reanalysis data sets are in close 15 agreement with each other, with differences from the observations of between 0.5 K and 1 K. 16 The altitude and pressure of the cold point tropopause are captured best by ERA5, CFSR,   (Fig. 7). To show differences at relatively high spatial resolution, we focus on the   There is also evidence of a secondary maximum in the differences over equatorial South 5 America or the East Pacific, although the magnitude and location of this maximum differ among 6 the reanalyses. 7 The convective centre over the Western Pacific warm pool, where the cold point tropopause is 8 coldest, does not show enhanced biases relative to the observations. One possible explanation 9 for the bias distribution might link the enhanced temperature differences to Kelvin wave activity  The zonal mean lapse rate tropopause (Fig. 9) at the equator is found at similar temperatures than those based on radio occultation data. The latitudinal structure of lapse rate tropopause 1 temperatures reveals slightly larger biases at the equator and better agreement between 10°-20° 2 in each hemisphere, and is generally very similar to the latitudinal distribution of biases in cold 3 point temperatures (Fig. 6). The altitude of the lapse rate tropopause shows considerable zonal 4 variability, ranging from 14.5 km to 16.7 km. All reanalyses capture the plateau in lapse rate 5 tropopause altitudes between 20°S and 20°N and the steep gradients in these altitudes on the 6 poleward edges of the tropics.    . Both radiosonde records suggest significant cooling at the 70 hPa level (Fig. 12). 21 Trends derived from reanalysis data can be problematic due to changes in the assimilated 22 observations. Given this potential limitation, it is of interest to examine whether the reanalysis 23 trends are consistent with the hypothetically more reliable trends derived from homogenized 24 observational records. At 70 hPa, temperature trends based on the reanalysis data sets span 25 almost exactly the same range (-0.5 to -1.1 K/decade) as those based on the radiosonde data sets (-0.5 to -1 K/decade). All reanalysis-and observationally-based trends are significant at 1 this level, confirming the stratospheric cooling reported by many previous studies (e.g., Randel     Temperatures at the cold point and lapse rate tropopause levels show warm biases in reanalyses 22 when compared to observations. This tropopause-level warm bias is opposite to the cold bias 23 found at all model levels and is most likely related to difficulties in determining the true cold 24 point and lapse rate tropopause levels from discrete temperature profiles with coarse vertical 25 resolution. Our analysis confirms that the magnitude of the bias shift is consistent with the 26 vertical resolution of the reanalysis data, with the smallest bias shifts found for data sets with 27 the highest vertical resolution around the tropopause (ERA5 and CFSR). The negative 28 temperature bias at model levels is often cancelled out by the positive bias introduced when 29 identifying the lapse rate and cold point tropopause locations. As a result, ERA5, which has a 30 small negative bias at model levels, has the most realistic tropopause temperatures, while CFSR, 31 which produces the most realistic model-level temperature profile, has a warm bias of 0.6-0.9 32 K at the cold point and lapse rate tropopause levels. Older reanalyses like MERRA, JRA-25 33 and especially NCEP-NCAR R1 show the largest temperature biases at the tropopause levels.

34
The zonal structure of tropopause temperature reveals that the biases in reanalysis relative to 35 observations maximise at or near the equator. All of the recent reanalyses produce a realistic 36 horizontal structure of cold point temperature with minima corresponding to the centres of 37 tropical deep convection. Differences between reanalyses and observations are greatest over 38 equatorial Africa. These enhanced differences are possibly related to Kelvin wave activity and 39 associated disturbances in TTL temperatures that also maximize in this region.  Long-term reanalysis trends in temperature at 70 hPa show good agreement with trends derived 13 from adjusted radiosonde data sets. All reanalyses and observational data sets indicate 14 significant stratospheric cooling at this level of around -0.5 K/decade to -1 K/decade. At the 15 100 hPa and cold point levels, both adjusted radiosonde data sets and reanalyses indicate large 16 uncertainties in temperature trends. Reanalysis-based estimates at the cold point range from no 17 trend at all (0 K/decade for ERA-Interim) to strong cooling of -1.3 K/decade (NCEP-NCAR 18 R1). While the latter is outside of the observational uncertainty range and can be considered 19 unrealistic, all other reanalyses data sets agree with at least one of the observational data sets 20 within uncertainties. The bulk of the reanalyses are in good agreement at these levels, 21 suggesting small but significant cooling trends of -0.3 K/decade to -0.6 K/decade that are 22 statistically consistent with trends based on the adjusted radiosonde data sets. 23 Advances of the reanalysis and observational systems over the last decades have led to a clear 24 improvement of the TTL reanalyses products over time. In particular, the modern reanalyses 25 ERA-Interim, ERA5, MERRA2, CFSR and JRA-55 show a very good agreement after 2002 in 26 terms of the vertical TTL temperature profile, meridional tropopause structure and interannual 27 variability. Temperatures at the cold point and lapse rate, on the other hand, are too high for 28 most old and modern reanalyses. As these differences maximise over Central Africa, a centre 29 of deep convective activity, chemistry-transport models driven by reanalyses and simulating air 30 mass transport into the stratosphere, can be expected to have too little dehydrations and too high 31 water vapor. Depending on the particular application, different reanalyses offer different 32 advantages such as a realistic cold point temperature (e.g., ERA5), small bias in the TTL 33 temperature profile (e.g., CFSR), realistic spatial distribution of the cold point temperature (e.g.,