Introduction
Water vapour in the stratosphere plays a fundamental role in the radiative budget and affects the ozone chemistry in this
atmospheric layer. In the lower stratosphere water vapour is the most important greenhouse gas. As such, it is part of an
important global warming feedback mechanism. A warmer climate increases lower stratospheric water vapour, leading to an
even warmer climate. estimated this feedback to be 0.3 Wm-2 for a temperature anomaly of
1 K at 500 hPa. In addition, water vapour is a fundamental component of polar stratospheric clouds. The
heterogeneous chemistry on cloud particle surfaces is responsible for the severe ozone depletion in the lower stratosphere
during winter and spring, especially in the Antarctic . Water vapour is also the main source of
hydrogen radicals (HOx = OH, H, HO2) in the stratosphere that contribute to ozone destruction
through catalytic loss cycles .
Thus, any change of stratospheric water vapour over a longer timescale has important implications
e.g.. In the past, the
majority of studies related to longer-term water vapour changes were based on observations by the balloon-borne NOAA frost
point hygrometer (FPH) at Boulder (a more detailed description of the measurement principle is provided in
Sect. ). These observations have been performed since 1980, typically once per month, providing the
longest time series of water vapour in the lower stratosphere. Positive trends over Boulder were first reported by
, then by and , and finally . For the time period
from 1980 to 2010, showed an overall increase of 0.24–0.42 ppmvdecade-1 for the altitude
range between 16 and 26 km accompanied by significant variability on shorter timescales. A total of 25 % of the observed
increase could be associated with changes of methane . The oxidation of this trace gas is the most
important in situ source of water vapour in the stratosphere. The other relevant source of water vapour in the
stratosphere is transport from the troposphere, which mainly occurs through the cold tropical tropopause region. One major
pathway is slow ascent (accompanied by large horizontal motions; ) in which the amount of water vapour
entering the stratosphere is mainly controlled by the tropopause temperature (or better cold point
temperature;
). Different changes of this temperature have been reported. reported an
overall negative trend for the time period from 1980 to 2003, which would correspondingly result in a decrease in lower stratospheric water vapour. Recent work by indicates zero or slightly positive trends at the tropical
tropopause for the time periods 1979–1997 and 1998–2014. The other pathway thought to be of importance is the convective
lofting of ice particles . Once the particles reach the stratosphere, they
evaporate and enhance the amount of stratospheric water vapour. This process is not dependent on the (cold point)
temperature. Balloon-borne observations indicated no trend of the convective ice lofting into the stratosphere for the
time period between 1991 and 2007 . Based on all these results it is difficult to assess what
process(es) caused the 30-year net increase in lower stratospheric water vapour observed by the FPH observations at
Boulder .
(a) In green the approximated trend estimates derived from
the merged satellite data set for the latitude band between 35 and
45∘ N are shown. See text for more details. Below the dashed line
these estimates consider the time period from 1988 to 2010; above they are for
the time period from 1986 to 2010. In red and blue the corresponding trend
estimates derived from the FPH observations at Boulder are shown. The error
bars represent the 2σ uncertainty. The right axis provides an
approximation of the altitude in geometrical terms. This information is
derived from the MIPAS data. (b) Difference among the trend
estimates derived from the FPH observations and merged satellite data set.
![]()
Satellite observations of stratospheric water vapour exist since 1978 , with some gaps. The instruments
have limited lifetimes and thus individual data sets do not allow a trend analysis on the same timescale as the FPH
observations at Boulder. Recently, merged zonal mean data sets from seven satellite instruments. This
merging was achieved with the help of a CMAM simulation with specified dynamics (aka nudging), which acted as a transfer
function. For each data set biases relative to the CMAM simulation were estimated. This assumes that the CMAM simulation
provides a realistic representation of the water vapour variability (including trends) and that the satellite data sets do
not have a drift in the bias estimation period. With this bias information the individual data sets were then adjusted
relative to the Aura/MLS observations. Finally, the average over all bias-corrected data sets was used for the merged data
set. This data set covers the time period between 1986 or 1988 (depending on latitude and altitude) and 2010, providing
the opportunity to evaluate the trends observed by the FPH observations at Boulder and to address water vapour changes on
a more global scale. The trends derived from the merged satellite data set for the zonal mean of the latitude around
Boulder were negative below about 10 hPa and positive above. This behaviour could also be essentially observed at
all other latitudes. Below 20 hPa the percentage changes up to 2010 were typically between -10 and -5 %,
which roughly corresponds to a trend between -0.2 and -0.1 ppmvdecade-1. attributed
this trend to a reduced transport of water vapour into the stratosphere as a consequence of lower tropopause temperatures
and a changed circulation in the stratosphere. During the same period as covered by the merged satellite data set, the FPH
observations at Boulder still exhibit a clear increase in lower stratospheric water vapour .
Figure provides a summary of the trend discrepancies between the FPH observations and the merged
satellite data. The trends derived from the merged satellite data set for the latitude band around Boulder
(35–45∘ N) are shown in green. Below (above) the dashed line the satellite trends are representative for the
time period from 1988 to 2010 (1986–2010). The estimates are based on a digitisation of Fig. 5a in . The
extracted percentage trends were converted to volume mixing ratio trends using an average profile derived from all
Aura/MLS observations in the latitude between 35 and 45∘ N from August 2004 to December 2010. These data are
chosen as all other satellite data sets are finally adjusted to the MLS data in the merging of , as
described above. Accordingly, the trends presented in Fig. are approximations. About the
uncertainty of these trends we only know that they are at least statistically significant at the 2σ uncertainty
level. Since the actual uncertainty level is unknown to us, we conservatively assume that the uncertainty is exactly at
the 2σ level (which certainly overestimates the trend uncertainties). In red (blue) the trends derived from the FPH
observations at Boulder are given for the time period from 1986 to 2010 (1988–2010). These were obtained by means of
multilinear regressions (see Eq. later). Only small differences are observed between the two time
periods. The trend estimates do not change significantly if the vertical resolution of the FPH data is adjusted to that of
the satellite observations. Likewise smoothing the FPH observations in time (with a 1-year running average), to reduce the
scatter among individual observations, does not notably affect the trend estimates.
Figure b shows an estimate of the trend differences between the
FPH observations and the merged zonal mean satellite data set. The differences vary with altitude ranging from about 0.3
to 0.5 ppmvdecade-1. Given the importance of water vapour in the lower stratosphere there is a dire need to
reconcile these differences. Potential explanations could be the following or a combination of
these.
There might be problems with one of the data sets or even with both.
The location of Boulder might be not representative for the zonal mean due to local processes specific for the
location, for example (American monsoon, lee of the Rocky Mountains, etc.)
There might be unresolved differences among the measurement techniques, like due to the different spatial and
temporal sampling and resolution.
In their discussion of the trend discrepancies between the FPH observations and the merged satellite data set
opted for the second possible explanation, indicating that the temporal behaviour at Boulder is
different than for the zonal mean of the latitude band around the Boulder latitude. Trends derived from the CMAM
simulation at 100 hPa (considering the time period 1980–2010) indicated longitudinal differences at
40∘ N, but also at other latitudes. Subsampling the simulation to Boulder yielded better correlations with the
FPH observations, in particular with respect to interannual variations. However, the trends derived from the FPH observations
and the model simulation still disagreed, even in sign.
In this study we compare trend estimates for the Boulder location and the zonal mean for the latitude band around the
Boulder latitude considering multiple time periods. For that we use several model simulations and observational data
sets. The observations are meant to study this aspect on a decadal scale while the simulations will be used to analyse
even longer time periods. This aims to understand how large the trend differences are in general and how much they might
contribute to the trend discrepancies shown in Fig. b. In the next section the
model simulations and observational data sets are briefly described while Sect. outlines the
analysis approach. The results of our analysis are presented in Sect. and subsequently discussed in
Sect. .
Approach
Boulder time series
For the Boulder time series we consider simulated data and satellite observations that are spatially located
within.
a 1000 km radius around the Boulder FPH observation site.
the latitude band between 35 and 45∘ N.
In the analysis of the HALOE data set we use less strict criteria because of its sparseness relative to the other data
sets. Instead of the radius criterion, data in the wider longitude range between 130 and 80∘ W are considered.
In temporal terms we consider two sets of data for the Boulder time series. Set no. 1 simply comprises all data in a given
month. We will refer to these time series as full time series. Set no. 2 is adapted to the individual FPH observations at
Boulder. From that we can also assess the role of the temporal sampling for the trend differences. For the simulations the
data from the closest time step are used. For the observations all data obtained within ±12 h of the FPH measurements
are considered. We will refer to these time series as adapted time series.
All data obeying the spatial and temporal criteria are combined to monthly means. For the observations we consider only
monthly means that are based on at least five measurements to avoid spurious results. As a result, a temporal adaption to the
individual FPH observations is only meaningful for the MLS observations.
Zonal mean time series
For the zonal mean time series we consider monthly means of all data in the latitude range between 35 and 45∘ N,
resembling the merged satellite data set. Monthly zonal means derived from the satellite observations are discarded if
they are not based on a minimum number of 20 measurements. If a monthly mean does not exist for the Boulder time series,
e.g. because there were no FPH observations for the adapted time series or due to screening of the satellite data, this
monthly mean is also not considered for the zonal mean results.
In addition, we also investigate how the trend estimates at Boulder compare to those for zonal means of other latitude
bands. For that we consider the latitude bands 45–55∘ N, 25–35∘ N, Equator–60∘ N and
60∘ S–60∘ N. The first two bands are adjacent to the latitude band around the Boulder latitude. The
last two bands cover a wider range of latitudes. This aims to investigate how representative trends at Boulder are on
regional and more global scales.
De-seasonalised time series for Boulder (a), the
35–45∘ N zonal mean (b) and its difference (c)
for a number of model simulations considering the pressure level of
70 hPa. Results labelled with the suffix (A) are adapted to the
actual FPH observations at Boulder; see text for more details. The time ticks
consider the middle of the specified years.
![]()
De-seasonalisation
In our analysis we employ de-seasonalised data. This enhances the visibility of the long-term behaviour and has the
positive side effect that the MIPAS observations from the FR and RR periods are homogenised. Between these periods
typically a small bias in the absolute water vapour volume mixing ratios exists. The de-seasonalisation is achieved by
means of regression, again motivated by the MIPAS data. This approach has the advantage of working for time series that
cover a time period between 12 and 24 months, which applies here to the MIPAS data for the FR period. The regression model
contains an offset and a parametrisation for the semi-annual (SAO) and annual variation (AO) using orthogonal sine and
cosine functions:
fd(t,ϕ,z)=Coffset(ϕ,z)+CSAO1(ϕ,z)⋅sin(2⋅π⋅t/pSAO)+CSAO2(ϕ,z)⋅cos(2⋅π⋅t/pSAO)+CAO1(ϕ,z)⋅sin(2⋅π⋅t/pAO)+CAO2(ϕ,z)⋅cos(2⋅π⋅t/pAO).
In the equation fd(t,ϕ,z) denotes the fit of the regressed time series for a given time t (in years),
latitude band ϕ and altitude z, which is subsequently subtracted from the absolute time series to obtain the
de-seasonalised time series. C are the regression coefficients of the individual model components and pSAO
and pAO represent the time periods of the SAO (0.5 years) and AO (1 year),
respectively. The regression coefficients are derived according to the method outlined by , using
the standard errors of the monthly means (their inverse squared) as statistical weights. Autocorrelation effects and
empirical errors are not considered in this regression.
For the de-seasonalisation of the simulations we consider data in the time period from 1985 to 2010. The start year is
chosen because of obvious differences in the water vapour abundances among the simulations, related to differences in
their initial conditions and spin-up time (see Fig. and
Sect. ). The last year that is covered by all simulations is 2010. For the observations it
is not possible to use a consistent time period. Instead the entire time period covered by the individual data sets is
used for the de-seasonalisation.
Trend estimates and trend differences
Like the de-seasonalisation of the time series, the estimation of the water vapour trends is based on regression. For this analysis the regression model is as follows:
ft(t,ϕ,z)=Coffset(ϕ,z)+Ctrend(ϕ,z)⋅t+CSAO1(ϕ,z)⋅sin(2⋅π⋅t/pSAO)+CSAO2(ϕ,z)⋅cos(2⋅π⋅t/pSAO)+CAO1(ϕ,z)⋅sin(2⋅π⋅t/pAO)+CAO2(ϕ,z)⋅cos(2⋅π⋅t/pAO)+CQBO1(ϕ,z)⋅QBO1(t)+CQBO2(ϕ,z)⋅QBO2(t).
In comparison to the regression model used for the de-seasonalisation, it contains, in addition, a trend term
Ctrend and a parametrisation for the quasi-biennial oscillation (QBO). In our analysis we determine only
a single trend for the entire time period. Trend changes within this period are correspondingly not analysed (see
e.g.). Even though the regression is applied to de-seasonalised time series the SAO and AO terms
are kept since the regression models for the de-seasonalisation and trend analysis differ. The QBO parametrisation is
based on normalised winds at 50 hPa (QBO1) and 30 hPa (QBO2) observed over Singapore
(1∘ N, 104∘ E), which are almost orthogonal. These data are provided by Freie Universität Berlin
(webpage: http://www.geo.fu-berlin.de/met/ag/strat/produkte/qbo/qbo.dat, last access: 6 June 2018).
Unlike for the de-seasonalisation, in this
regression we consider autocorrelation effects and empirical errors to obtain optimal estimates of
the trends and their uncertainties.
To be consistent with our motivation shown in Fig. , we calculate the water vapour trends
separately for the Boulder time series and the zonal mean time series and subsequently derive the trend
differences. Correspondingly, the trend differences (ΔCtrend) and their uncertainties
(εtrend) are given as
ΔCtrend(ϕ,z)=CtrendBoulder(z)-Ctrendzonal mean(ϕ,z)εtrend(ϕ,z)=εtrendBoulder(z)2+εtrendzonal mean(ϕ,z)2.
Here CtrendBoulder represents the trends derived from the Boulder time series and
εtrendBoulder are the corresponding uncertainties.
Likewise,
Ctrendzonal mean and εtrendzonal mean denote the trends calculated
from the zonal mean time series and their uncertainties.
Results
In this section we will first present the simulation results and subsequently the results derived from the
observations. We focus on the altitude range between 100 and 20 hPa that is typically covered by the FPH
observations and in almost all cases completely in the stratosphere .
Simulations
Figure shows the de-seasonalised Boulder time series
(Fig. a) and the zonal mean time series around the Boulder latitude (latitude range 35–45∘ N, Fig. b)
at 70 hPa for the different model simulations. The differences between the two time series are shown in Fig. c as a complement. The time
series adapted to the individual FPH observations (see Sect.
and ) are marked with the suffix (A) in the figure legend. Overall, the Boulder and
the zonal mean time series are visually rather similar, with the latter being more smooth. The difference time series show
occasionally larger deviations (up to 0.6 ppmv in absolute terms); however any conspicuous behaviour or a trend
appears to be absent. In general, the different simulations yield similar results for Boulder and the zonal mean. The most
prominent exception is observed in the early 1980s. This relates to differences in the 1979 initial conditions and the
spin-up times among the simulations. Until 1985 the EMAC anomalies are significantly lower than for the other
simulations. In the first years the largest anomalies are found in the CMAM results, which were probably caused by higher
water vapour volume mixing ratios in the initial conditions based on the nudging of ERA-40 data (see
Sect. ). Presumably the best representation is provided by CLaMS, which, as a Lagrangian model, does not
need to deal with these aspects.
Trend estimates for the different model simulations for
Boulder (a), the zonal mean for the latitude band between 35 and
45∘ N (b), and their corresponding differences (c).
The different rows consider different time periods as indicated in the title
of the centre panels. Trends and trend differences significant at the
2σ uncertainty level are marked by triangles.
![]()
Figure shows the trend estimates for the time series at
Boulder (Fig. a) and the zonal mean for the latitude band between 35 and 45∘ N
(Fig. b). Figure c shows the corresponding difference according to
Eq. (). The different rows consider different time periods, i.e. 1985–2010 (top row),
1990–2010 (middle row) and 1995–2010 (bottom row). This is also indicated in the title of the centre panels. We have not
included the time period from 1980 to 2010 here. The differences in the water vapour anomalies among the simulations in
the early 1980s primarily affect the trends for Boulder and the zonal mean, yet the trend differences are comparable to
those for the time period from 1985 to 2010. Trends and trend differences significant at the 2σ uncertainty levels
are marked by triangles.
For the time period between 1985 and 2010, the EMAC results exhibit positive trend estimates at Boulder. They range
between 0.04 and 0.12 ppmvdecade-1. The results derived from the other simulations indicate rather small
trends at Boulder, typically within ±0.05 ppmvdecade-1. Small quantitative differences exist between the
results derived from the full and the adapted time series. Those from the adapted time series are typically larger (up to
about 0.04 ppmvdecade-1). Overall, the spread among the trend estimates ranges from about
0.1 ppmvdecade-1 at 100 hPa to 0.16 ppmvdecade-1 at 20 hPa. The trend estimates
derived from the zonal mean time series for the latitude band between 35 and 45∘ N look very similar to those
derived for Boulder. Correspondingly, the trend differences between Boulder and the zonal mean are very small. The
differences never exceed 0.04 ppmvdecade-1 in absolute terms. The largest differences are derived for EMAC
and WACCM (based on the adapted time series) at 100 hPa. The trend differences are predominantly positive below
70 hPa and mostly negative above 50 hPa. The exact altitude dependence differs in details among the
different simulation results.
Both for Boulder and the zonal mean, the trend estimates for the time period from 1990 to 2010 are negative. There are
differences among the individual simulations. The agreement is, however, better than for the time period between 1985 and
2010. The spread maximises at 100 hPa with about 0.12 ppmvdecade-1 and is smallest above
40 hPa with about 0.06 ppmvdecade-1. Differences among the results derived from the full and adapted
time series are very small. The trend differences between Boulder and the zonal mean are of a similar size as for the time
period from 1985 to 2010. A similar behaviour in terms of the altitude dependence is also visible.
The last time period we consider is from 1995 to 2010. At 100 hPa consistently positive trends are found, except
for the adapted EMAC time series. Overall, the trend estimates vary between -0.02 and
0.14 ppmvdecade-1. With increasing altitude the trend estimates typically decrease and above 45 hPa
they are all negative. Higher up, the trends continue to become more negative, except in the CMAM results. At 20 hPa
the trend estimates vary between -0.24 and -0.08 ppmvdecade-1 among the simulations, with significantly
smaller differences between the results derived from the full and the adapted time series. The best agreement among the
simulations is observed around 80 hPa where the spread is about 0.08 ppmvdecade-1. The altitude
dependence and the spread among the simulations is similar for the trend estimates derived from the zonal mean time
series. Quantitatively there are larger differences between the Boulder and zonal mean trends, clearly surpassing those
observed for the other time periods. Above 60 hPa the differences are still within
±0.02 ppmvdecade-1. Below this altitude the differences occasionally exceed
±0.05 ppmvdecade-1. The largest trend differences are derived from the adapted EMAC and CMAM time series.
The temporal development of the trend differences between the
Boulder and the zonal mean (35–45∘ N) time series, based on 11-year
time intervals. The results are given at the centre of the corresponding time
intervals, i.e. in 1995 for the time period between 1990 and 2000. The black
lines indicate zero trend differences.
![]()
To expand on the temporal development of the trend differences between Boulder and the zonal mean even more we derive
these differences continuously for 11-year periods, as shown in Fig. . The
results are assigned to the centre of the considered period, e.g. to 1995 for the time period between 1990 and 2000. The
trend differences vary with time and altitude in size and sign. On this shorter timescale the differences are typically
larger than observed for the longer time periods described in the last figure. There is also a more prominent distinction
between the results derived from the full and the adapted time series. The latter yield larger differences on an absolute
scale, but also some patterns are different.
For the full time series the trend differences are generally within ±0.04 ppmvdecade-1. Exceptions from
this behaviour are primarily observed at the lowermost altitudes. In particular the EMAC results exhibit significantly
larger differences, increasing to about ±0.15 ppmvdecade-1 at 100 hPa. The temporal development
of the trend differences exhibits a number of common features among the simulations, even though quantitative differences
are obvious. At the lowermost altitudes all simulations show negative trend differences from 1990 to about
1999. Afterwards positive trend differences are found. Higher up, i.e. above about 50 hPa, positive trend
differences are visible from 1995 to about 2004.
The trend differences derived from the adapted time series are within ±0.08 ppmvdecade-1 above
60 hPa. Below, they increase again in absolute size, maximising at about 0.2 ppmvdecade-1. The
different simulations agree on some difference patterns, as observed for the results derived from the full time
series. Most prominently, above 50 hPa the trend differences are typically positive from about 1990 to 2003 and
negative afterwards. The bisection of trend differences at the lowermost altitudes derived from the full time series is
only visible in some simulations. Finally, it should be noted that none of the trend differences shown in
Fig. are statistically significant at the 2σ uncertainty level.
Comparison of the trend estimates at Boulder (a) and the
zonal means for different latitude bands (b) as indicated in the
title. As in Fig. the (c) panels
show the difference between the two trends. The comparisons consider the time
period between 1987 and 2010. The (a) panels are all the same and are
repeated for convenience. Trends and trend differences significant at the
2σ uncertainty level are again marked by triangles.
![]()
To investigate the representativeness of the Boulder trends on a larger geographical scale
Fig. compares them to zonal mean trends for five latitude bands, namely
35–45∘ N (row no. 1), 45–55∘ N (row no. 2), 25–35∘ N (row no. 3), Equator–60∘ N
(row no. 4) and 60∘ S–60∘ N (row no. 5). The figure considers the time period between 1987 and 2010,
approximately the time coverage of the merged zonal mean satellite data set. The results in the left column are the same
for all rows and kept for the sake of convenience. The trends at Boulder are close to those obtained for the time periods
1985–2010 and 1990–2010 shown in Fig. . Note that in
Fig. the x axis is smaller, allowing a more detailed picture.
Overall, the trends at Boulder are within ±0.07 ppmvdecade-1. Clear differences among the simulations
exist, while the differences between the results derived from the full and the adapted time series are typically
smaller. The trend estimates derived from the full time series are again larger than those determined from the adapted
time series, with few exceptions. The EMAC results indicate positive trends (up to almost
0.1 ppmvdecade-1). Statistical significance is only visible at the highest altitudes. The trends derived from
the CLaMS data are negative below 35 hPa and positive above, ranging from about -0.05 to
0.05 ppmvdecade-1. They are relatively constant up to 60 hPa before they start to increase
significantly. The WACCM and CMAM trends show a similar altitude dependence with maximum negative trends (around
-0.05 ppmvdecade-1) in the altitude range between 50 and 40 hPa. For WACCM the trend estimates
become positive below 80 hPa while those derived from the CMAM simulation are negative at all altitudes.
As observed in Fig. the trends derived from the zonal mean time series for the
latitude band between 35 and 45∘ N are very similar to those for Boulder. Correspondingly, the trend differences
are small, i.e. ranging from -0.02 to 0.04 ppmvdecade-1. The differences are typically positive below
70 hPa and mostly negative above, affirming the picture observed for the time periods 1985–2010 and 1990–2010 in
Fig. .
As Fig. but here showing several observational results.
![]()
The trends derived from the zonal mean time series for the other latitude bands exhibit many common features with the
results for the zonal mean between 35 and 45∘ N. There are quantitative changes, but overall the trend estimates
remain of the same order. In addition, the altitude dependence of the trends also remains very similar and so do the
relations among the different simulations. The trend differences between Boulder and the zonal mean for 45 and
55∘ N remain within ±0.04 ppmvdecade-1. Again, the differences are typically positive below
70 hPa and predominantly negative at higher altitudes. The trend differences between Boulder and the zonal means
for 25 to 35∘ N and from the Equator to 60∘ N are quite similar, at least up to about 35 hPa. In
both cases the trend differences are within ±0.03 ppmvdecade-1. Typically the EMAC and CLaMS results are
at the higher end of this interval while the WACCM and CMAM results are at the lower end. The largest trend differences
compared to Boulder are observed for the zonal mean of the latitude band between 60∘ S and 60∘ N. These
range from -0.04 ppmvdecade-1 to slightly more than 0.06 ppmvdecade-1. There is clear
separation between the CLaMS results and those from the other simulations. For the CLaMS simulation the trend differences
are negative at 100 hPa (around -0.015 ppmvdecade-1). Around 90 hPa they turn positive and
continue to increase within increasing altitude. At 20 hPa the differences amount to
0.05 ppmvdecade-1 for the adapted time series and 0.06 ppmvdecade-1 for the full time series,
respectively. The other simulations indicate positive trend differences at 100 hPa. Around 70 hPa the
differences become negative and peak in absolute size between 50 and 40 hPa (between -0.04 and
-0.02 ppmvdecade-1). Higher up, the trend differences become less negative again.
Observations
Figure shows the de-seasonalised Boulder time
series (Fig. a) and the zonal mean time series around the Boulder latitude
(Fig. b) at 70 hPa for the HALOE, MIPAS and MLS observations. In
Fig. c the differences between the two time series are again shown, as previously in Fig. c. For
MLS there is also a data set that is adapted to the FPH observations at Boulder (see
Sect. ). Like the simulations the observations exhibit a rather similar picture for
Boulder and the zonal mean. The difference time series occasionally indicate some larger deviations. For example, in the
second half of 2011 some substantial positive differences are observed, consistent in the MIPAS and MLS data. The largest
differences typically occur in the MLS data set that is adapted to the FPH observations and for the HALOE data set,
primarily due to its sparseness. In addition, there is a notable agreement between the MIPAS and MLS time series for
Boulder and the zonal mean time series.
As Fig. but here again for the observations.
![]()
Figure compares the trend estimates at Boulder with those derived from zonal mean
time series for various latitude bands. The results for the different observational data sets consider different time
periods as indicated in the figure legend. Thus, they are not comparable and will be addressed separately.
In the lower stratosphere the HALOE observations exhibit negative trends at Boulder for the time period between 1992 and
2005. This behaviour is primarily related to the significant drop in lower stratospheric water vapour in 2001
. The relative dryness continued until 2005 (coinciding with the end of the
HALOE observations), causing the 14-year HALOE trends to be negative. The largest trends are observed below 80 hPa
with values around -0.45 ppmvdecade-1. Above, the trends become less negative with increasing altitude. At
20 hPa the trend amounts to about -0.03 ppmvdecade-1 (and is not statistically significant). The
trends derived for the zonal mean between 35 to 45∘ N have a similar altitude dependence, but their absolute
sizes are smaller. Accordingly, the trend differences between Boulder and the zonal mean are negative. Above
80 hPa the differences are almost invariant with altitude. Here they amount to about
-0.05 ppmvdecade-1. At lower altitudes the differences are larger, maximising at 100 hPa with about
-0.1 ppmvdecade-1. For the other latitude bands the zonal mean trends exhibit the same kind of altitude
dependence as observed for the band from 35 to 45∘ N. The most prominent variation concerns the exact altitude
at
which the negative trends exhibit their absolute maximum. For the latitude band between 45 and 55∘ N this occurs
close to 90 hPa. For the trends derived from the zonal mean from the Equator to 60∘ N and from
60∘ S to 60∘ N this maximum is observed around 70 hPa. The trend differences between Boulder and
the zonal mean for the latitude band between 45 and 55∘ N range from -0.1 to 0 ppmvdecade-1 with
the largest absolute values occurring below 75 hPa. For the latitude band between 25 and 35∘ N the trend
differences are close to zero, except below 75 hPa where they become significantly more negative. For the
remaining two latitude bands the trend differences are quite similar. At 100 hPa the trend differences amount to
-0.15 ppmvdecade-1. Towards 60 hPa the differences increase to around
0.05 ppmvdecade-1. Between 60 and 30 hPa the trend differences are rather constant. Higher up, they
increase to more than 0.1 ppmvdecade-1.
The MIPAS observations indicate positive trends at Boulder during the time period from 2002 to 2012. The trends decrease
with increasing altitude from about 0.25 ppmvdecade-1 at 100 hPa to 0.1 ppmvdecade-1 at
20 hPa. For the zonal mean between 35 and 45∘ N the trend estimates are also consistently
positive. However, they show a slightly different altitude dependence than for Boulder. Below about 70 hPa the
trends increase while higher up they decrease. In correspondence, the trend differences between the Boulder and zonal mean
estimates are most pronounced below about 70 hPa, rising to 0.05 ppmvdecade-1 at
100 hPa. Above 70 hPa the differences are close to zero. A very similar behaviour is observed for the
trend differences between Boulder and the zonal mean considering the latitude band between 45 and 55∘ N. The
trend differences to the estimates for the latitude bands from 25 to 35∘ N and the Equator to 60∘ N
exhibit a pronounced altitude dependence. They decrease from more than 0.1 ppmvdecade-1 at 100 hPa to
-0.05 ppmvdecade-1 at 20 hPa. The sign of the trend differences switches at about
60 hPa. The trend differences between Boulder and the zonal mean for 60∘ S to 60∘ N are positive
at all altitudes. The smallest differences are close to zero and are observed between 45 and 30 hPa. The largest
difference is visible at 100 hPa with 0.15 ppmvdecade-1.
The Boulder trends derived from the MLS observations from 2004 to 2016 are positive. They exhibit a pronounced altitude
dependence. The trend estimates exhibit maxima at 70 hPa (0.4 ppmvdecade-1) and 30 hPa (close
to 0.3 ppmvdecade-1). Minima are found at 100 (0.2 ppmvdecade-1), 45 and
20 hPa (around 0.25 ppmvdecade-1). The trends derived from the adapted time series are slightly
larger than those calculated from the full time series. The trend differences between these two data sets are of a similar
order as observed for the simulations addressed before. The MLS trends derived from the zonal mean time series for the
different latitudes indicate a similar altitude dependence to that observed for Boulder. Overall, the trend differences
between Boulder and the zonal means are generally within ±0.05 ppmvdecade-1. Prominent exceptions occur
below 70 hPa for the differences to the zonal means from 25 to 35∘ N, Equator to 60∘ N and
60∘ S to 60∘ N. Here, the differences can be as large as 0.15 ppmvdecade-1. In addition,
the trend differences between Boulder and the zonal mean from 60∘ S to 60∘ N are noticeably larger than
for the other latitude bands, ranging from 0.05 to 0.15 ppmvdecade-1. Beyond that, the trend differences are
consistently larger (by about 0.05 ppmvdecade-1) for the adapted time series at altitudes around
40 hPa.
Discussion and conclusions
In this work we compared trend estimates for lower stratospheric water vapour between Boulder and zonal mean data around
the Boulder latitude (35 to 45∘ N) considering different time periods. For that we analysed multiple data sets,
both from simulations and observations. The objective was to quantify how large these trend differences typically are and
how much they could possibly help to explain the discrepancies in the trend estimates between the FPH observations at
Boulder and a merged zonal mean satellite data set . For the time period from the
late 1980s to 2010 the trend differences (FPH minus merged zonal mean satellite data set) range from 0.3 to
0.5 ppmvdecade-1, increasing with altitude.
Our analysis shows that there are differences in the trend estimates between Boulder and the zonal mean, both for the
simulations and observations. These trend differences are dependent on altitude and the time period considered.
For the time period from the late 1980s to 2010 the simulations indicate trend differences between about -0.02
and
0.04 ppmvdecade-1 (which are however not statistically significantly different from zero). These are clearly
smaller than the discrepancies in the trend estimates derived from the FPH observations and the merged satellite data
set. The larger positive differences are observed close to 100 hPa. Here, the trend differences partly
resolve
the observational discrepancies. Above about 60 hPa the trend differences derived from the model simulations are
however typically negative. This indicates that the trend estimates for the zonal mean data should be larger than at
Boulder, which is contradictory to the observed trend differences between the FPH observations and the merged zonal mean
satellite data set. Also, the simulations do not exhibit any pronounced deviations in the trend differences derived from
time series using all data during a given month (which we referred to as full time series) or just using that closest in
time to the actual FPH observations (which we referred to as adapted time series). This indicates that the temporal
sampling has only a small influence on the trend differences on this timescale.
Given these model results, a different temporal behaviour between Boulder and the zonal mean is not a viable explanation
for the discrepancies in the trend estimates derived from the local FPH observations and the merged zonal mean satellite
data set presented by . It still could be the case that the simulations underrepresent variability or
that the trend differences originate from smaller spatial and temporal scales than are resolved by the model simulations
(i.e. sub-grid processes). For the Boulder time series we used data in a 1000 km radius around the Boulder FPH
observation site and within the latitude range from 35 to 45∘ N. These criteria were primarily chosen
for consistency with the analysis of the satellite observations whose exact measurement locations vary from orbit to orbit
and day to day. In an additional analysis of the simulations, we considered for the Boulder time series only data from the
closest grid point in space (EMAC: 40.5∘ N, 104.1∘ W, Δr = 109 km;
WACCM: 40.7∘ N, 105.0∘ W, Δr = 80 km; CMAM: 39.0∘ N, 105.0∘ W,
Δr = 113 km; CLaMS: 40.0∘ N, 105.0∘ W, Δr = 17 km). This
analysis yields small quantitative changes (not shown here). Qualitatively, exactly the same conclusions can be drawn as
from the standard analysis. The temporal resolutions of the analysed simulations vary (see
Sect. ). The CMAM simulation provides the best resolution in this analysis with 6 h. Accordingly
the worst temporal mismatch to the actual FPH observations is 3 h. This gives an upper limit of temporal scales not
covered in this analysis. However, arguably the different simulations yield similar results, as do the analyses of the full
and the adapted time series.
For a single decade of data the trend differences between Boulder and the 35–45∘ N zonal mean are typically
larger than those discussed above for the entire time period from the late 1980s to 2010. The differences are typically
within ±0.10 ppmvdecade-1, except close to 100 hPa where the differences can be occasionally as
large as ±0.2 ppmvdecade-1. For the simulations, the trend differences derived from the adapted time
series are typically larger than the trend differences obtained from the full time series on an absolute scale. A factor of 2
is a common feature. In the MLS data, significant trend differences between the full and the adapted time series are
observed around 40 hPa. These differences should be kept in mind when comparing results for Boulder and the
zonal mean on the shorter timescales.
In addition, we analysed trend differences between Boulder and the zonal means for a number of latitude bands. This aimed
to investigate how representative the Boulder trends are for a more global scale. For the time period from the late 1980s
to 2010 the simulations indicate trend differences within the interval from -0.04 to 0.06 ppmvdecade-1. The
largest differences occur when the Boulder trends are compared to those for the zonal mean of the latitude band between
60∘ S and 60∘ N. Based on these results, the Boulder trends should be quite representative (or
a reasonable first guess) for the trends on more global scales during this time period. The caveats regarding missing
variability or sub-grid processes in the simulations apply here as well. For shorter time periods, as covered by the
individual satellite data sets, the representativeness becomes smaller in general.
From our analysis it appears that a continued search for the reasons of the trend discrepancies between the FPH
observations at Boulder and the merged satellite set is necessary (see list in the Introduction). In addition, even more
differences become apparent. To start with, this considers the simulations themselves. The overall spread among the trend
estimates derived from the different simulations can be almost as large as 0.2 ppmvdecade-1. For the time
period from the late 1980s to 2010 the spread varies between 0.06 and 0.12 ppmvdecade-1. To some degree this
relativises the trend discrepancies between the FPH observations and the merged zonal mean satellite data set, if the
spread among different simulations amounts to a considerable fraction of the discrepancies themselves. The reasons for the
spread among the simulations are probably manifold, comprising general model characteristics (e.g. parameterisations, wave
forcing, convection scheme), the choice of the nudged reanalysis data (and their quality over
time;
) or the exact details of the nudging (e.g. parameters, top height, relaxation time; see
Sect. ). Our analysis does not provide clear hints in a specific direction but leaves room for
obvious followup activities.
Then, the trend estimates obtained from the simulations also differ from those derived from the FPH observations and the
merged satellite data set (compare Figs. and ). Overall,
they are closer to the trend estimates from the merged satellite data set, but consistently larger by about 0.05 to
0.2 ppmvdecade-1 depending on simulation and altitude. Compared to the FPH trend estimates the model results
are consistently smaller by about 0.1 to 0.45 ppmvdecade-1. In many ways this situation is reminiscent of the
results presented by that indicated clear trend differences among the FPH observations, HALOE and
a simulation from an older version of WACCM for the time period between 1992 and 2002.
Differences between the de-seasonalised time series obtained from
the FPH observations and the Boulder time series derived from the different
simulations and observational results at 70 hPa. In addition, the
differences between the FPH time series and the zonal mean (35 to
45∘ N) time series for EMAC and SAGE II are shown. To provide
a clearer picture, the differences are smoothed with a 1-year running
average. At least three data points are required for a valid running average.
The dashed-dotted line indicates the time period covered by the merged
satellite data set at this altitude. The time ticks again consider the middle
of the specified years.
![]()
A way forward is certainly to put more focus on understanding differences in time series of the water vapour anomalies
instead of those in derived quantities. An example of this is shown in Fig. , which considers
the difference between the de-seasonalised time series derived from the FPH observations at Boulder and the Boulder time
series from the different simulations and satellite observations (i.e. FPH minus the other data sets) used in this work at
70 hPa. For the simulations and the MLS data set the adapted Boulder time series are used while for the HALOE and
MIPAS data sets the full time series are employed, as also indicated in the legend of the figure. In addition, the
difference time series between the FPH observations and the zonal mean (35 to 45∘ N)
results exemplarily
from the EMAC simulation as well as the ERBS/SAGE II (Earth Radiation Budget Satellite/Stratospheric Aerosol
and Gas Experiment II) observations, derived with retrieval version 7.00 , are shown. These differences
are marked with the suffix (zonal) in the figure legend. The ERBS/SAGE II data are considered here as they are also part
of the merged satellite data set and actually define the start of the time series . The
de-seasonalisation period of the FPH time series is always adapted to the time series to which it is compared, i.e. from 1985 to
2010 for the model simulations, from 1992 to 2005 for HALOE, from 2002 to 2012 for MIPAS, from 2004 to 2016 for MLS and
from 1988 to 2005 for the SAGE II data set (see Sect. ). For a clearer picture the
differences are smoothed with a 1-year running mean. At least three valid data points during this period are required for
a running mean to be considered further. The differences visible in the figure are also representative for other
altitudes, even though some details are different. A number of aspects gain
attention.
The differences of the EMAC Boulder and zonal mean time series from the FPH observations are very similar. This is
also true for the other simulations (not shown here). It highlights once more the main outcome of this study that the
temporal behaviour at Boulder largely resembles that for the zonal mean around the Boulder latitude. More obvious
deviations occur in the EMAC simulation between 1997 and 2000. This behaviour is also found in the WACCM simulation and to
some degree in the CLaMS results, while the CMAM simulation shows larger deviations around 1990. Also, in 2004 pronounced
deviations are observed, consistently in all simulations (see also Fig. ).
Before 1986 the differences from the FPH observations are predominantly negative (EMAC being the exception), while
afterwards until 2011 they are mostly positive. As the trends in this work are derived using multilinear regression with
a single trend term, this behaviour is consistent with larger trend estimates for the FPH observations compared to the
simulations for the time period from the late 1980s to 2010.
While the SAGE II differences from the FPH observations mostly blend with the other data sets there is pronounced
deviation between 1989 and 1991 (afterwards data are screened due to aerosol contamination by the Pinatubo
eruption). During this time period the differences are more negative than for the model simulations. This behaviour is
consistently observed below 30 hPa. Since this is close to the very beginning of the merged time series it has
a pronounced effect on the trend estimates. It provides an explanation of why the trend estimates derived from the merged
satellite data set are smaller than those for the simulations considering the time period from the late 1980s to
2010. Overall, this might hint at a potential issue with the SAGE II data before the Pinatubo eruption. Alternatively, an
issue might originate from the equal weighting of the pre- and post-Pinatubo SAGE II data in the merged satellite data
set. More investigations are required to rule out any of these potential issues.
The temporal development of the differences is quite consistent in qualitative terms for the various simulations and
observational data sets. Features like the strong negative differences around 1993/94, the subsequent increase until 2000,
the relatively constant behaviour from 2001 to 2009 or the decrease starting in 2010 are visible for all simulations and
satellite observations. Interestingly, we also find a similar behaviour in difference time series between frost point
hygrometer observations at other stations and the simulations and satellite observations used in this work (not shown
here). Explicitly, this applies to the NOAA FPH observations at Lauder (45∘ S, 169.7∘ E) and the CFH
(cryogenic frost point hygrometer; ) observations at San Jose (9.9∘ N, 84.0∘ W) and
Lindenberg (52.2∘ N, 14.1∘ E). In quantitative terms, the consistency of the differences is evidently less
good. The spread among the various data sets is on average 0.26 ppmv and is thus comparable to the
differences between the FPH observations and the different simulations and satellite observations themselves. In
particular, between 1980 and 1985 there are huge deviations among the simulations in their differences from the FPH
observations, relating to differences in the initial conditions and the spin-up times among the simulations (except for
CLaMS). After this period the average spread decreases to 0.21 ppmv.
In summary, understanding the differences shown in Fig. and their temporal development,
hopefully in combination with the merged satellite data set, should be a focal point of further research on lower stratospheric water vapour. This will inevitably yield better consistency in the trend estimates but also highlight the
benefit of combining different data sources, such as in situ observations, satellite measurements and modelling efforts.