Reliable long-term data records are necessary to detect and understand climate change. Vertically resolved trends in temperature provide a sensitive test of the mechanisms driving climate change which in turn are vital to improving climate model projections of future climate. The National Oceanic and Atmospheric Administration (NOAA) operated Microwave Sounding Units (MSUs) on polar-orbiting satellite platforms from 1978 to 2005 and Advanced Microwave Sounding Units (AMSUs) from 1998 onwards. These nadir-looking instruments became the primary source of long-term temperature climate records of the troposphere and stratosphere even though their measurements were originally intended for short-term weather forecasts, not climate applications. To create long-term homogeneous climate data records, measurements need to be adjusted to account for inter-satellite biases, orbital changes and calibration deficiencies, as well as small differences in radiometer frequency and bandwidth between the MSU and AMSU instruments. Three such long-term records are currently maintained by independent groups – Remote Sensing Systems (RSS) , the University of Alabama at Hunstville (UAH) , and the Centre for Satellite Applications and Research (STAR) . While substantial differences in temperature trends were found between the RSS and UAH data sets in the stratosphere , trends in the lower stratosphere between (A)MSU data sets are generally comparable .

Satellite-based infrared sounders, as well as Global Positioning System (GPS) radio occultation (RO) experiments, offer high vertical resolution making them prime candidates to extend the MSU temperature record and improve the monitoring of upper-air temperature changes. Temperature measurements from infrared sounders typically have a measurement uncertainty of <2 K with good horizontal resolution. RO measurements have moderate horizontal resolution (∼300 km), but good vertical resolution (∼1 km) and are well suited for high quality climatologies because of their low systematic errors (<0.5 K) and high stability . found good agreement between RO and radiosonde temperature anomalies, but statistically significant differences in trend estimates from (A)MSU data compared to RO measurements which they believed resulted mainly from the (A)MSU data.

In this study, monthly mean zonal mean (5∘ zones) temperature time series from six satellite-based instruments, including three RO instruments, are merged to create a new time series of monthly mean temperatures on 17 pressure levels from 300 hPa to 7 hPa. This data set is then vertically integrated using the (A)MSU vertical weighting functions to create a (A)MSU proxy record suitable for assessing the continuity of the merged (A)MSU temperature data sets by the RSS and UAH groups, focusing on the Temperature Lower Stratosphere (TLS) product. The TLS series spans altitudes from 13 to 22 km and results from merging the MSU channel 4 (MSU4) and AMSU channel 9 (AMSU9) temperature time series.

Section  describes the data sets used to compile our new vertically resolved temperature data set. In Sect. , the method used to merge the six source data sets is described. The merged data set is then validated against RATPAC (Radiosonde Atmospheric Temperature Products for Assessing Climate) radiosonde data, COSMIC (Constellation Observing System for Meteorology, Ionosphere, and Climate) RO data as well as the NCEPCFSR (National Centers for Environmental Prediction Climate Forecast System Reanalysis) reanalyses in Sect. . In Sect. , the new data set is vertically integrated using an appropriate weighting function to create an (A)MSU proxy. The resulting temperature time series is then used to verify the merging of MSU4 and AMSU9 temperature series.

Monthly mean zonal mean temperature records from satellite missions by the European Space Agency (ESA) and Third Party Missions (TPMs), as well as RO data, were created in collaboration with SPARC (Stratosphere–troposphere Processes And their Role in Climate) as a project of SPIN (ESA SPARC Initiative). In the following sections, we briefly describe the characteristics of the different instruments, and how the monthly mean zonal mean data sets used in this study were produced from the individual temperature profiles.

The monthly mean zonal mean temperature data sets from the three ESA/ESA-TPM instruments and the three RO experiments described in Sect.  were merged to create new zonal mean monthly mean temperature time series in 5∘ latitude zones, henceforth referred to as the vertically resolved temperature (VRT) climate record or data set. As not all of the six source data sets had measurements at all 27 pressure levels mentioned in Sect. , only the first 17 pressure levels were used. RO measurements have been shown to be an effective climate benchmark below ∼25 km altitude . Comparisons of the annual mean zonal mean temperatures calculated from the three RO data sets show agreement within 0.5 K in the lower stratosphere (∼20 km) but there is a robust bias that increases in magnitude with height, reaching 3–5 K in the upper stratosphere, near 40 km. GRACE is consistently warmer than both CHAMP and TSX in the Southern Hemisphere and consistently colder in the Northern Hemisphere. Because of the systematic bias of GRACE against both CHAMP and TSX, we exclude GRACE as a suitable benchmark data set for the construction of the merged product. Choosing between CHAMP and TSX as an initial data set, CHAMP provides a longer data record.

Figure  shows graphically the process of how, starting with CHAMP as the benchmark data set, the remaining data sets are successively merged. At each latitude zone and pressure level, the data set with the biggest temporal overlap with CHAMP is chosen as the first candidate to be merged with CHAMP. In the example of Fig. , MIPAS is found to have the biggest overlap with CHAMP. Thus, differences are calculated between the CHAMP and MIPAS monthly mean temperatures and these are then modelled statistically using a regression model that includes an offset and drift where the offset fit coefficient is expanded in Fourier pairs to also account for seasonality in the differences, i. e.: f(t)=a1+a2t+a3sin⁡(2πt)+a4cos⁡(2πt)+a5sin⁡(4πt)+a6cos⁡(4πt). The uncertainty of the difference function f(t) takes into account possible correlations between the coefficients σf2(t)=∑i=16∂f(t)∂ai2⋅σai2+2∑i=15∑k=i+16∂f(t)∂ai∂f(t)∂ak⋅cov(ai,ak) The statistical model fit from Eq. () is then used to correct the MIPAS data TMIPAS,corr(t)=TMIPAS(t)+f(t), and the model's uncertainty is used to recalculate the uncertainty estimate of the corrected MIPAS data σMIPAS,corr(t)=σMIPAS2(t)+σf2(t), where σMIPAS is the uncertainty on the MIPAS mean temperatures. A new benchmark time series is created by merging the original CHAMP time series with the corrected MIPAS time series by calculating a weighted average Tmerged=wCHAMP⋅TCHAMP+wMIPAS,corr⋅TMIPAS,corrwCHAMP+wMIPAS,corr, where the weights are given by wx=1/σx2. The uncertainty in the correction applied to the MIPAS monthly means is incorporated into a new estimate of the uncertainty on the monthly mean σmerged=1wCHAMP+wMIPAS,corr The process is then repeated using each remaining data set in the order of decreasing temporal overlap with the iteratively revised benchmark data set. For the example in Fig. , after merging MIPAS with CHAMP, this is found to be GRACE. Differences between the GRACE monthly mean temperatures and those of the new benchmark data set of CHAMP merged with corrected MIPAS are then used as input to a new statistical model to derive corrections to be applied to the GRACE monthly means. This process is repeated, consecutively folding the MIPAS, GRACE, TSX, ACE-FTS and SMR into the original CHAMP data to successively extend the CHAMP data set. The order in which the data sets are merged can be different for each latitude zone and pressure level depending on the availability of data.

The inclusion of each additional data set reduces the uncertainty on the final product (as long as they are appropriately corrected). The monthly mean data from an instrument are excluded from this merging process if there are fewer than 4 measurements in a particular month or less than 5 % of measurements of the month with the highest number of measurements within that year. The new VRT climate record consists of such a temperature time series for every 5∘ latitude zone and on the following 17 pressure levels: 300, 250, 200, 170, 150, 130, 115, 100, 90, 80, 70, 50, 30, 20, 15, 10, and 7 hPa.

The merged temperature database described above was validated against the following three independent data sets:

The radiosonde-based RATPAC-A  database.

One of the purposes of constructing the VRT data record was to assess the quality of the merging of the MSU4 and AMSU9 temperature time series in the lower stratosphere (the TLS product) by both the RSS and UAH group. Of particular interest is whether the transition from the period of overlap between MSU4 and AMSU9 (from 1998 to 2006) to coverage by AMSU9 only after 2006 introduces any discontinuity in the climate data record. A statistical model is used that accounts for any systematic biases between the VRT time series and the MSU4+AMSU9 data sets. The residuals are examined for any anomalies, in particular a step function around the transition to the AMSU9 only measurements around 2006.

A new database of monthly mean zonal mean temperatures spanning the upper troposphere and stratosphere was created by merging monthly mean zonal mean temperature measurements from the mid-infrared spectrometers ACE-FTS and MIPAS, the microwave sounder SMR, and the RO experiments GRACE, CHAMP and TSX. The new temperature data record is aggregated in 5∘ latitudinal zones and is vertically resolved on pressure levels from 300 to 7 hPa.

Systematic biases between different instruments were corrected by statistically modelling the differences allowing for an offset and drift as well as temporal periodic fluctuations. The merging process was initiated with the CHAMP temperature time series as RO measurements are known to provide an effective climate benchmark. As the merging algorithm mainly removes systematic biases relative to the initial data set, it is crucial to choose this initial benchmark data set carefully. The sensitivity to the initial benchmark data set was tested by using the same merging algorithm, but starting the process with MIPAS, GRACE and TSX, respectively, rather than CHAMP. A comparison of these latter three temperature time series to that initiated with CHAMP shows all resultant merged data sets being in very good agreement. The main differences between the different data sets is an offset of similar magnitude to the offset seen in the right hand column of Fig. . The merged temperature set starting with CHAMP typically lies somewhere in the middle of the three comparative data sets. However, for some latitude zones and pressure levels, especially in the Southern Hemisphere, the other three temperature time series show a trend relative to the one started with CHAMP. A trend in the differences is more problematic because one intended use of our merged data set is to detect relatively small warming or cooling trends in the atmosphere. This trend is relatively small for TSX, but bigger for MIPAS and GRACE. Nevertheless, the comparison shows that CHAMP remains the best choice as an initial benchmark data set in the merging process.

The merged VRT climate record was validated against three different databases. The first validation was done against RATPAC-A which provides annual mean temperature anomalies from radiosonde measurements aggregated over seven large geographical regions. The inter-annual variability in temperatures is well represented by our VRT data set over most of the upper troposphere and stratosphere at the ±0.5 K level, but can exceed 1 K at the 250 and 300 hPa levels. The comparison with the COSMIC RO database showed more divergent results. Temperature differences were typically up to 1 K at most pressure levels (≥50 hPa), but could reach up to 4 K at lower pressure levels. There were also clear spatial differences between tropical and extra-tropical zones depending on altitude. The temperatures from our VRT database were consistently lower in the tropics than in the extra-tropics above 200 hPa, but this pattern reversed below 200 hPa. The third validation was performed against NCEPCFSR reanalyses. Differences were typically less than 2 K over most pressure levels, but could reach 5 K at the highest altitudes (≤10 hPa). Seasonal variations, on the other hand, were more pronounced than in the other two validations.

As a major application, the VRT temperature record was used to verify the quality of the merging process of the MSU4 and AMSU9 channels of both the RSS and UAH groups. After removing systematic biases, the residuals relative to our iVRT data set were examined for statistically significant break-points. No statistically significant steps were found in the residuals around the switch from MSU4 to AMSU9, confirming that both groups made appropriate adjustments in the merging process to assure a continuous temperature time series that is not affected by calibration errors between the two types of instruments or other systematic biases due to satellite drifts. Only in the two polar 5∘ latitude zones did our iVRT temperature record show an increase in variability in 2005 relative to the UAH temperature series.

The VRT climate record can also be seen as a contribution to establish long-term vertically resolved temperature records to enable updated knowledge of long-term changes in temperatures across different altitude ranges. To date, the microwave sounding measurements have been a de-facto standard for temperature climate data records. With the increased use of satellite infrared-sounders as well as highly accurate RO measurements, more finely vertically-resolved data records are available. Our VRT record is, to the best of our knowledge, the first that merged temperature measurements from selected ESA/ESA-TPM missions with RO measurements. It has been shown that the uncertainty on the monthly mean zonal mean temperatures decreases with an increased number of instruments used in the merging. The VRT climate record thus provides a useful data set for long-term temperature trend analyses.