We define the spread as the difference between the maximum and minimum volume mixing ratio among the data sets at a given time and place. As such, the spread is a simple measure of the collective consistency among the time series from the different data sets. We have chosen this approach for the spread calculation since for the other approaches based on standard deviation or percentiles, assumptions have to be made. However, we have also calculated the spread using the other two approaches and derived qualitatively the same results as for the maximum–minimum calculation. Prior to the spread calculation, we performed an additional screening among the data sets to avoid unrepresentative spread estimates. The screening is again based on the median and median absolute difference, as done before for the monthly zonal mean calculation. Monthly zonal means outside the interval 〈median[yp(t,ϕ,z)i]±7.5 MAD[yp(t,ϕ,z)i]〉 were not considered, with i=1,…,nd(t,ϕ,z) and nd(t,ϕ,z) denoting the number of data sets at a given time, latitude and altitude. The subscript p is used as a placeholder either for the absolute or the de-seasonalised data. This screening removed overall 2.6 % of the data for the latitude band between 80∘ and 70∘ S. For the tropical and the mid-latitude bands, respectively 3.6 % and 3.7 % of the data were removed. Subsequently, the spread was derived. We did not impose any additional criterion on the number of data sets available for a spread estimate to be valid (two data sets is the natural minimum). However, for much of the 1990s the only available satellite data sets are HALOE and SAGE II. Since both instruments provide solar occultation measurements, the number of coincidences is limited. Thus, their time series do not constantly overlap, there are many gaps in the spread. Therefore, we focus in the results section on the time period between 2000 and 2014.

To describe the consistency between two time series we employed the correlation coefficient r(ϕ,z): r(ϕ,z)=∑i=1nt(ϕ,z)yp(ti,ϕ,z)1-yp‾(ϕ,z)1⋅yp(ti,ϕ,z)2-yp‾(ϕ,z)2∑i=1nt(ϕ,z)yp(ti,ϕ,z)1-yp‾(ϕ,z)12⋅∑i=1nt(ϕ,z)yp(ti,ϕ,z)2-yp‾(ϕ,z)22, with yp‾(ϕ,z)1=1nt(ϕ,z)∑i=1nt(ϕ,z)yp(ti,ϕ,z)1,yp‾(ϕ,z)2=1nt(ϕ,z)∑i=1nt(ϕ,z)yp(ti,ϕ,z)2.

The subscripts at the end of the variables refer to the two data sets. p is again a placeholder for the absolute and de-seasonalised data. nt(ϕ,z) is the number of months the two time series actually overlap, i.e. where both data sets yield valid monthly means. Correlation coefficients were only considered if the overlap was at least 12 months. We did not perform any significance analysis for the coefficients since we simply want to show if the expected high correlation between two time series exist.

Figure  shows the de-seasonalised water vapour time series for the southern polar latitudes. The HIRDLS, SCIAMACHY (solar occultation) and SAGE III observations have no coverage in this latitude region, while the GOMOS observations' coverage is too limited to allow derivation of de-seasonalised time series. In the de-seasonalised time series, a spread among the data sets can be found at the four altitudes considered in the comparison. The largest anomalies and the largest spread are found at 0.1 hPa (up to ±2 ppmv), while the smallest anomalies and thus the smallest spread is found at 3 hPa (generally in the range of ±0.4 ppmv).

At 0.1 hPa the time series start from 1991 onwards with HALOE, since SAGE II measurements are not available at this altitude. Large differences in the seasonal variation of the de-seasonalised time series are found, resulting in a considerable spread among the data sets, larger than at other altitudes. Large anomalies (up to ±2 ppmv), and thus large inter-annual variation, are found for the MIPAS-Oxford V5H, MIPAS-ESA V5R and MIPAS-ESA V7R data sets, while quite small anomalies are found for both ACE-FTS data sets. These large anomalies in the above mentioned MIPAS data sets are a consequence of the pronounced (spiky) seasonal variation in the absolute data (see Fig. S1 in the Supplement) that is difficult to be accounted for in the sinusoidal regression used for the de-seasonalisation.

Decadal changes in water vapour are found in the de-seasonalised time series at 3 hPa. Several periods of water vapour increases are followed by water vapour decreases. Negative anomalies are found around 1992 while positive anomalies are found around 1996 (HALOE). Water vapour then shows positive anomalies again in ∼2003 (HALOE, POAM III, SAGE II), followed by a decrease in 2003–2004, which again is followed by a slight increase in water vapour that lasts until 2010. From 2010 onwards water vapour remains unchanged. The last increase in water vapour is most strongly pronounced in SMR 489 GHz indicating a drift in the SMR 489 GHz data relative to the other data sets (see also Sect. ). A large spread between the de-seasonalised time series is found between 1999 and 2004 (mainly between POAM III, SAGE II and SMR 489 GHz). Between 2005 and 2014, good agreement between the de-seasonalised time series is found. However, SMR 489 GHz has somewhat higher anomalies (from 2011 onwards) than the other satellite data sets.

At 10 hPa, the spread among the data sets is quite similar to that observed at 3 hPa, but the variability in water vapour is more pronounced. There is a decrease in the SAGE II de-seasonalised water vapour time series of 1986–1990. An increase in the de-seasonalised water vapour time series is found in POAM III around 2001. Also from 2009 onwards there seems to be a slight increase in water vapour in all data sets. The SMR 489 GHz de-seasonalised time series at 10 hPa is in good agreement with the de-seasonalised time series of the water vapour products derived from the other satellite instruments. However, the SMR 489 GHz as well as the SOFIE anomalies are low relative to MLS. This becomes quite obvious at the end of the time series (2012–2014), when only ACE-FTS, MLS, SMR 489 GHz and SOFIE were taking measurements. Also, the influence of the QBO is clearly visible at this altitude level. Distinct positive anomalies are found in 2007–2008, 2011 and 2013.

At 80 hPa the water vapour distribution is strongly influenced by dehydration (Sect. ). The de-seasonalised time series at 80 hPa once again depict the spread between the individual instruments in this latitude band. At 80 hPa similar results as for 10 hPa are derived (except that here no long-term changes are visible). However, here the deviations between HALOE and SAGE II are smaller than at 10 and 3 hPa. As at 10 hPa, a decrease in the anomalies of the SAGE II de-seasonalised time series is found for 1986–1990. The de-seasonalised time series then remains constant until 1998 (HALOE and SAGE II). From 1998 onwards the spread between the data sets increases. There is an increase in the anomalies found in 2001, which is followed by a decrease, which lasts until 2004. Another decrease in water vapour is found in 2009. At 80 hPa, POAM III shows stronger inter-annual variation and higher/lower anomalies than at 10 and 3 hPa, depending on which year is considered.

Figure  shows the de-seasonalised water vapour time series for the tropics. The POAM III, SAGE III, SCIAMACHY (solar and lunar occultation) and SOFIE data sets have no coverage in this latitude band. In the SAGE II time series some data gaps occur which are due to the aftermath of the Pinatubo eruption (resulting in unrealistically high water vapour values that were filtered out) as well as the “short events” between June 1993 and April 1994, when too few measurements were available . In the tropics, good consistency between the data sets is found except at 0.1 hPa, where again the spread between the data sets is largest. At 0.1 hPa some data sets exhibit larger anomalies (±1.2 ppmv; e.g. MIPAS-Oxford V5H and MIPAS-ESA V7R), while others exhibit rather small anomalies (±0.3 ppmv; e.g. ACE-FTS and MLS). The HIRDLS, GOMOS and MAESTRO (80 hPa) data sets show generally larger anomalies and thus larger spread than the other satellite data sets. The de-seasonalised time series in the tropics reflect the decadal changes in water vapour that have been documented in the literature, such as the drop in stratospheric water vapour after 2000 and in 2012 . Further, at 3 and 10 hPa, a variability in water vapour on an approximate 2-year timescale associated with the QBO is clearly visible.

At 0.1 hPa the time series starts in 1991 with the HALOE data set, which is also the only one available for these altitude and latitude regions until 2001. The de-seasonalised time series from HALOE shows an increase between 1992 and 1996 followed by a period with rather constant anomalies that lasts until 2001. Afterwards a decrease is visible until 2005. SMR 489 GHz observes, in contrast to HALOE, an increase in water vapour between 2001 and 2005. Therefore, at the beginning of the SMR 489 GHz record the anomalies at 0.1 hPa are clearly lower than those from HALOE or the other satellite data sets measuring from 2001 onwards. However, a large spread between the data sets is also found during this time period. A similar increase (but somewhat stronger) is found in the MIPAS Oxford V5H data set between 2001 and 2003, but here the anomalies are higher than the ones from the other satellite data sets. While the MIPAS Oxford V5H and SMR 489 GHz data sets show increasing anomalies, the other data sets show decreasing anomalies. From 2006 onwards all data sets show increasing anomalies. Between 2012 and 2014, ACE-FTS, MLS and SMR 489 GHz are the only data sets covering this time period and deviations among them are quite visible. SMR 489 GHz anomalies are higher and show larger inter-annual variability than ACE-FTS and MLS. MLS (together with ACE-FTS) exhibit generally the lowest anomalies (±0.3 ppmv) compared to the other satellite data sets at this altitude.

At 3 and 10 hPa the time series begins with SAGE II in 1986. From 1991 onwards HALOE observations are also available. Both SAGE II and HALOE provide here a much better representation of the temporal development of the water vapour time series and the inter-annual variability than in the Antarctic since both data sets have a much better temporal coverage in the tropics (see Figs. S1 and S2 in the Supplement). SAGE II shows somewhat larger anomalies than HALOE. Generally, the de-seasonalised time series show good agreement with each other at these two altitude levels (3 and 10 hPa). Further, at these altitude levels, the lowest anomalies and the lowest spread between the data sets is found, especially at 10 hPa. The deviations between MLS (or ACE-FTS) and SMR 489 GHz found during the time period 2012–2014 are still evident at 3 hPa but to a much lesser extent than at 0.1 hPa. At 3 hPa, inter-annual variations (with anomalies roughly on the order of ±1 ppmv) due to the QBO are clearly visible. At 10 hPa this variability is far less obvious. Also, the differences between SMR 489 GHz and the other data sets measuring during the time period 2001–2005 (SAGE II and HALOE) are found to a lesser extent at 3 hPa, but not at 10 hPa. The GOMOS data set exhibits large scatter. At 10 hPa the HIRDLS data set indicates stronger inter-annual variability than the other satellite instruments. This level is the uppermost altitude where HIRDLS can be retrieved and accordingly the data here are more uncertain. Both drops in water vapour, the one in 2001 and the one in 2012, are clearly visible in the de-seasonalised time series at 10 hPa. The latter one is strongly pronounced in the three remaining data sets covering that time period (ACE-FTS v3.5, MLS and SMR 489 GHz). There is also a clear variability on an approximate 2-year timescale associated with the QBO visible at this altitude level, although not at all times are as clearly pronounced as at 3 hPa.

Similar to the other three pressure levels, at 80 hPa relatively good agreement between SAGE II and HALOE is found. However, SAGE II typically shows somewhat lower anomalies than HALOE. At 80 hPa, higher variability with larger anomalies than at 10 and 3 hPa is found (generally around ±0.8 ppmv). The data sets agree well in terms of the inter-annual variation. The drops in 2000 and 2011 are consistently observed, as are the recoveries afterwards. This is also true for the pronounced QBO in 2006–2008. In 2005 the MIPAS-Bologna V5R NOM and MIPAS-ESA V5R NOM data sets show strong negative anomalies (up to -2 ppmv) which are not found in the other data sets. Similar behaviour of these data sets is found in 2011, when they show strong positive anomalies (up to 1.6 ppmv), while in the other satellite data sets, anomalies up to only 0.4–0.8 ppmv are found. MAESTRO shows strong scatter, mainly because 80 hPa is near the upper altitude limit of the MAESTRO water vapour retrieval. Another distinctive characteristic in the de-seasonalised time series at 80 hPa is the increase in water vapour that lasts until mid-2014 (ACE-FTS v3.5, MLS and SMR 544 GHz) which is anti-correlated with the time series at 10 hPa.

Figure  shows the de-seasonalised time series for the Northern Hemisphere mid-latitudes. The GOMOS, SCIAMACHY lunar and SOFIE data sets have no coverage in this latitude region. As for the other latitude bands the largest spread between the satellite data sets is found at 0.1 hPa. This is accompanied by large inter-annual variability. The ACE-FTS v3.5, MIPAS-Bologna V5H, MIPAS-Oxford V5H and SMR 489 GHz data sets are among the data sets showing the largest inter-annual variability and also the largest anomalies at 0.1 hPa. The MIPAS-Oxford V5H data set covers the time period of 2002–2004 and here the largest anomalies (exceeding 2 ppmv) are found. The largest negative anomalies are found in 2005 and 2006 with -1.6 and -2 ppmv, respectively. The differences between ACE-FTS v3.5 and the other satellite data sets become most pronounced at the end of the data record when only SMR 489 GHz and MLS were still measuring. Here, ACE-FTS v3.5 shows some larger variability. At this altitude, the drift in the SMR 489 GHz data set is again visible. The anomalies are typically more negative compared to the other data sets until 2004, while they are more positive after 2012. The HALOE data set indicates an increase in water vapour until about 1997 and a decrease afterwards. There appears to be a decrease in water vapour for all data sets from 2007 to 2010, followed by a pronounced increase that lasts until early 2012.

At 3 hPa, the de-seasonalised time series show generally good agreement, while at 10 hPa the best agreement is found. Differences at 3 hPa are that SMR 489 GHz exhibits lower anomalies during the time period 2001 to 2006 and higher anomalies than the other data sets from 2010 to 2014 and that SAGE II shows higher anomalies than the other satellite instruments at the end of their data record (2004–2005). Differences at 10 hPa are found in the time period 2004–2008, when SAGE II and HIRDLS show stronger inter-annual variability, and during 2010–2012, when SMR 489 GHz exhibits somewhat higher anomalies than the other satellite data sets. In both altitude levels, an increase in water vapour between 1992 and 2000 (10 hPa) and 1992 and 1998 (3 hPa) is found. The two water vapour drops that occurred after 2000 and in 2011 in the tropics are also visible at 10 hPa in the Northern Hemisphere mid-latitudes, however with a temporal delay.

Although the inter-annual and decadal variability at 80 hPa is low, some satellite data sets (MAESTRO, POAM III and SMR 544 GHz) show larger deviations from the other satellite data sets. In the MAESTRO data, high inter-annual variability is found with anomalies reaching up to 1.6 ppmv. In this altitude region, MAESTRO has its best temporal coverage in the mid-latitudes, but still 80 hPa is at the upper limit of the MAESTRO measurements and therefore not every measured profile reaches that high up. This explains why higher variability (scatter) than in the other satellite data sets is found for the MAESTRO time series. POAM III exhibits much larger anomalies than the other satellite data sets (+1.2 ppmv compared to ±0.4 ppmv). Although the POAM III anomalies decrease with time, they still remain higher than the anomalies from the other satellite data sets. The differences between POAM III and the other satellite data sets are caused by the limited temporal sampling (only summer months are measured) of POAM III in this latitude region making the de-seasonalisation by regression apparently fail. In the SMR 544 GHz data set, larger inter-annual variability is found, but with much smaller anomalies than MAESTRO. In the SAGE II data, the anomalies are decreasing slightly in the time period 1987–2002. Further, there is some pronounced QBO alongside an overall increase from 2004 to 2012.

Overall, in the Northern Hemisphere mid-latitudes, the lowest inter-annual variability is found, especially at 80 hPa. Similar to the comparisons in the Antarctic and tropics, the largest inter-annual and decadal variability as well as the largest spread between the data sets is found at 0.1 hPa. The drops in stratospheric water vapour after 2000 and in 2011 that are observed in the tropics are also found at 10 hPa in the mid-latitudes, but with a temporal delay and to a lesser extent than in the tropics.

Figure  shows the correlation between the de-seasonalised MIPAS-Oxford V5R NOM time series and those from the other data sets for the Antarctic, tropics and the Northern Hemisphere mid-latitudes. The largest spread in the correlation between the satellite data sets is found in the Antarctic (Fig.  top), also where the lowest correlation over all altitude levels is found (rarely exceeding a correlation coefficient of 0.8). MIPAS-ESA V5R NOM and MIPAS-ESA V7R are among the data sets showing the highest correlation with MIPAS-Oxford V5R NOM over all altitude levels while the lowest correlation with MIPAS-Oxford V5R NOM is found for SCIAMACHY lunar throughout most altitudes. The SOFIE and SMR 544 GHz data sets show very low correlations (even negative for SOFIE) at the lowest altitude levels (below 10 hPa) as well as above 3 hPa (but here SMR 489 GHz instead of SMR 544 GHz). In between these altitudes levels the SOFIE and SMR 489 GHz data sets show similar correlation to MIPAS-Oxford V5R NOM as the other data sets.

In the tropics (Fig.  middle), the correlation coefficients vary between 0.8 and 1 for most data sets between 30 and 1 hPa. Low correlations are found for all data sets between 100 and 30 hPa, except the MIPAS-IMKIAA V5R NOM data set, which shows a high correlation (>0.8) up to 1 hPa with MIPAS-Oxford V5R NOM. The data sets that show the lowest correlation with MIPAS-Oxford V5R NOM (even in some occasions negative) are GOMOS and MAESTRO. These data sets thus deviate from the typical correlation of most other data sets. Above 60 hPa and above 25 hPa this is also true for HIRDLS and SMR 544 GHz, respectively. These two data sets show reasonable correlation with MIPAS-Oxford V5R NOM at the lowest altitude levels, but then the correlation coefficients decrease rapidly with increasing altitude, most likely due to increased measurement noise. At altitudes above 0.7 hPa the correlation decreases for all data sets and the spread between the data sets increases. For MIPAS-ESA V5R NOM, the correlation, although decreasing, remains rather high with a correlation coefficient of 0.7. The lowest correlation at 0.1 hPa is found for the ACE-FTS v2.2, ACE-FTS v3.5, MIPAS Bologna V5R NOM and MIPAS-Bologna V5R MA data sets.

In the Northern Hemisphere mid-latitudes (Fig.  bottom), the correlation coefficients vary between 0.4 and almost 1 in the altitude region between 0.7 hPa and 10 hPa depending on which data set is considered. The spread in the Northern Hemisphere mid-latitudes is almost as large as the spread in the Antarctic. Very high correlation (correlation coefficient of around 0.9–1) between MIPAS-Oxford V5R NOM and the other data sets is found at, for example, around 1 hPa for the MIPAS-ESA V5R NOM and MIPAS-ESA V7R data sets. The lowest correlation between MIPAS-Oxford V5R NOM and the other data sets is found above 1 hPa for the two ACE-FTS data sets while the SMR 489 GHz data set shows a rather low correlation throughout the entire altitude region considered in this study. Below 10 hPa the lowest correlations (even negative correlations) are found for HIRDLS, MAESTRO, SCIAMACHY limb and SMR 544 GHz data sets. These data sets also deviate from the usual spread in correlation of the data sets.

The correlation of all data sets is given in Figs. – in form of matrix plots for the three latitude bands and four altitude levels. In addition to the correlation coefficient, the number of months of overlap between the time series is given (requiring a minimum of 12 months; see Sect. ). The same figures for the correlation of the absolute time series are given in the Supplement (Figs. S8–S10). The correlation matrix shown in Fig.  gives a good overview over the temporal consistency of all data sets in the Antarctic. The correlations between the data sets are generally positive (green), but in some cases negative correlations (red) are found, for example, in the case of the correlation between the MIPAS-IMKIAA V5H and POAM III data sets at 10 hPa or that between the MLS and SCIAMACHY lunar data sets at 3 hPa. However, in these two cases, the number of overlapping months is not that high (14 and 28) and this may explain the low correlation between these data sets. An example of where a negative correlation is found despite the high number of overlapping months (70) is the correlation between the MIPAS-Bologna V5R NOM and MLS data sets at 0.1 hPa. An example of a high number of overlapping months (114) and high correlation coefficient is the correlation between the MLS and SMR 489 GHz data sets at 10 hPa. Nevertheless, although in the Antarctic the correlation is generally positive, the correlation coefficient rarely exceeds 0.5. An exception is the 3 hPa level, where a generally high correlation among the MIPAS data sets is found. Similar behaviour between the MIPAS data sets is found at 10 hPa.

In Fig.  the correlation matrix for the tropics is shown. The large spread between the data sets we found in Fig.  at 0.1 hPa is also reflected in the correlations among all data sets. The same holds for the good correlations that are found at 3 and 10 hPa. An exception here is the GOMOS data set that shows negative correlations with all instruments at 3 hPa, but the number of overlapping months is rather low. At 80 hPa the spread between the data sets is not as large as at 0.1 hPa, but still larger than at 3 and 10 hPa. At 80 hPa occasionally negative correlations are found. This primarily concerns comparisons involving the GOMOS, HALOE, MAESTRO and MIPAS-Oxford V5H data sets. The lowest (negative) correlation is found between SMR 489 GHz and SAGE II data sets, but here the number of overlapping months (21) was also rather low.

The correlation matrix shown in Fig.  gives a good overview of the temporal consistency of all data sets in the mid-latitudes. The majority of the correlations are positive, but for some comparisons negative correlation is found. One such example is the correlation between the MIPAS Bologna V5H and SMR 489 GHz data sets at 3 hPa. However, again the number of overlapping months was rather low and may explain the negative correlation between these data sets. An example of negative correlation, despite a high number of overlapping months, is found between MIPAS-Bologna V5R NOM and MIPAS-Bologna-V5R-MA with MLS at 0.1 hPa. The correlation of these two data sets with the other data sets is also generally low at 0.1 hPa. Also, for the two ACE-FTS data sets the correlation of most data sets is often low despite a sufficient number of overlapping months. Positive correlations are found for the ACE-FTS v2.2/v3.5 data sets in comparison to the MIPAS-IMKIAA V5R MA, MIPAS-Oxford V5R MA, MLS and SMR 489 GHz. The highest correlation at 0.1 hPa is found between the two ACE-FTS data sets and between ACE-FTS v2.2 and MLS. At 3 and 10 hPa generally high correlations among the MIPAS data sets are found. At 10 hPa the correlation of HIRDLS with some data sets is high, but low with the other data sets. At 80 hPa low correlations between MAESTRO and all other instruments are found.

In summary, a high number of overlapping months does not necessarily guarantee a good correlation between two data sets, but generally the chances are quite high if this is the case. On the other hand, if data sets overlap only for a low number of months, good agreement between these data sets can still be found. Therefore, for assessing the agreement between two data sets, both the number of overlapping months and the correlation coefficient should be taken into account. The correlation assessment again confirms what we found before from the qualitative time series comparison, namely that the best agreement between the satellite data sets is found in the tropics, while in the Antarctic and Northern Hemisphere mid-latitudes a large spread between the data sets is found. Generally, the lowest correlations are found in the Antarctic. Further, in each latitude band the correlation is lower in the lower stratosphere and lower mesosphere than in the middle stratosphere.

Figure  shows that below 20 hPa large drifts (up to 2.5 ppmvdecade-1 and even higher) are found between SMR 489 GHz and the other satellite data sets. In the altitude region between 20 and 1 hPa, good consistency between the satellite data sets is found despite the different time periods of measurements. The smallest drifts, ranging from about 0 to 0.5 ppmvdecade-1, are found around 20 hPa. The drifts consistently increase with altitude and maximise around 0.4 hPa. Above 1 hPa the drifts of SMR 489 GHz vary between about 0.75 and 1.5 ppmvdecade-1 depending on which data set the SMR 489 GHz data set is compared to, but decrease with altitude towards 0.1 hPa. The drifts range here between 0 and 1.25 ppmvdecade-1. The drifts between SMR 489 GHz and the other satellite data sets are in most cases significant at the 2σ uncertainty level as can be seen from Fig.  (right panel). Larger drifts between SMR 489 GHz and the other data sets that obviously deviate from the majority of data sets are found for the comparison to the POAM III, SAGE II, SAGE III and HALOE data sets. However, this is due to the fact that for these data sets not only the overlap period with SMR 489 GHz is relatively short (4 years, 2001–2005), but also the number of months for which both data sets actually yield a valid monthly mean is small (see numbers given in figure legend). Additionally, these drifts are in most cases not statistically significant at the 2σ uncertainty level.

In Figs. – the drift estimates between the time series of all data sets are summarised as matrix plots for the three latitude bands and four altitudes. In the matrix plots, data sets are only shown if they yield any result at a given altitude. The drift estimates are based on the difference time series between the data sets given on the x axis and the data sets given on the y axis. Additional information that is given in the matrix plots includes the overlap period of the two data sets, how many months the two data sets actually overlap and if the drift is significant or not at the 2σ uncertainty level as well as the corresponding significance level for significant drift.

In the Antarctic (Fig. ), almost no significant drifts are found between the satellite data sets at the two lowest altitude levels (80 and 10 hPa). An exception here is the MAESTRO data set which shows a significant (negative) drift of -2 to -3 ppmvdecade-1 (significance level up to 3.7) and POAM III which shows a significant positive drift (2 to 3 ppmvdecade-1) compared to SAGE II and SMR 544 GHz (at 80 hPa). While the overall time period MAESTRO had overlap with other data sets was sufficiently long (>85 months), the number of coincident months for these data sets was rather low (9 months). Further, at 80 hPa, a significant negative drift is found between some MIPAS data sets and SOFIE. At 10 hPa, a significant (positive) drift (0.8 ppmvdecade-1) is found between the MIPAS-Oxford V5R NOM and ACE-FTS v2.2 data sets (significance level of 3.2) and of 2 ppmvdecade-1 between the SMR 489 GHz and POAM III data sets (significance level 3.0). Additionally, significant drifts are found between different MIPAS data sets relative to SMR 489 GHz and between the MLS and SMR data sets. At 3 hPa most drifts are significant. Most MIPAS data sets exhibit significant positive drifts relative to the ACE-FTS (significance level up to 5.7) and MLS (significance level up to 8.1) data sets. While in the comparisons to the ACE-FTS data sets the actual number of overlapping months is limited, this is not the case in the comparison to MLS. As before, for the SMR 489 GHz data set significant positive drifts are found (significance level up to 4.8) relative to most other data sets. A large variety of drifts is found at 0.1 hPa, but in most cases the drift is not significant. Data sets for which most drifts are significant at this altitude level are SMR 489 GHz (>2 ppmvdecade-1, significance level up to 6.4) and MIPAS-Bologna V5R MA (significance level up to 3.2).

In the tropics (Fig. ), larger drifts are found than in the Antarctic, especially at 0.1 hPa. Here, most drifts are significant. Significant drifts are found for the MIPAS-Bologna V5R NOM, MIPAS-Bologna V5R MA, MIPAS-ESA V5R, MIPAS-IMKIAA V5R NOM, MIPAS-Oxford V5R NOM and SMR 489 GHz data sets. For example, for MIPAS-Bologna V5R NOM and MIPAS-Bologna V5R MA drift (significance level up to 6.5) in comparison to most other satellite data sets is found. For MIPAS-Bologna V5R NOM this is also the case at 3 hPa (significance level up to 9.8). Large negative drifts are found for GOMOS (>-2.5 ppmvdecade-1, significance level up to 3.9) compared to most data sets. Also for SMR 489 GHz significant positive drifts (up to ∼1 ppmvdecade-1, significance level up to 8.5) for almost all data sets are found at 3 hPa. Good consistency is found among the MIPAS data sets. The drifts are low and in most cases not significant. An exception here is MIPAS-Oxford V5R NOM (∼0.6–1 ppmvdecade-1, significance level up to 9.8). For the tropics the best agreement among the data sets is found at 10 hPa. In most cases the drift is not significant and in cases where the drift is significant the drifts are relatively low with 0.2–0.4 ppmvdecade-1. Larger drifts are found at this altitude for GOMOS (up to -3 ppmvdecade-1) and HIRDLS (up to -2 ppmvdecade-1). For GOMOS the drifts are significant in most cases (significance level up to 4.3), while this is not the case for HIRDLS.

At 80 hPa a wide variety is found. Some data sets show positive drift, some negative. In some cases the drift is significant and in other cases not. For example, a positive drift (2 ppmvdecade-1) relative to almost all data sets is found for MIPAS-Bologna V5R NOM (significance level up to 6.4). For the HIRDLS data set a significant positive drift (also ∼2 ppmvdecade-1) is found compared to MIPAS-IMKIAA V5R NOM, MIPAS-IMKIAA-V5R MA and MIPAS-Oxford V5R NOM (significance level 2.0–4.6). A large drift (>3 ppmvdecade-1) at this altitude level is found for MIPAS-ESA V5R MA compared to MIPAS-IMKIAA V5R NOM (significance level 4.8). Also the MIPAS-Oxford V5R NOM shows significant drifts compared to a number of data sets.

The patterns of the estimated drifts in the Northern Hemisphere mid-latitudes shown in Fig.  are quite similar to the drifts in the tropics and Antarctica. However, the estimated change in ppmvdecade-1 seems to be somewhat lower in the mid-latitudes than in the tropics or Antarctic. The highest variety is again found at 0.1 hPa. Similar to the tropics significant drifts are found, e.g., for the MIPAS-Bologna V5R NOM and MIPAS-Bologna V5R MA (up to -2 ppmvdecade-1, significance level up to 3.9) data sets relative to the SMR 489 GHz data set. At 3 hPa, for most data sets the drifts are small and/or not significant. Significant negative drifts are found for both ACE-FTS data sets and for SMR 489 GHz. For SMR 489 GHz drift is found relative to most other data sets which is also in most cases significant. At 10 hPa HIRDLS shows pronounced drifts compared to the other data sets. However, these drifts are not significant except for the comparison with MLS (drift of 3 ppmvdecade-1, significance level 2.3). Otherwise for most data sets the drifts are small and/or not significant at 10 and 80 hPa. Exceptions are HIRDLS (-2 ppmvdecade-1) and MAESTRO (-1 ppmvdecade-1), which show negative drift at 80 hPa. For HIRDLS in most cases the drift is significant (significance level up to 4.1), but for MAESTRO in most cases not. For MIPAS-Bologna-V5R NOM significant positive drifts are found for all instruments that are in most cases around 0.2–0.4 ppmvdecade-1, but higher compared to HIRDLS (significance level 4.1), MAESTRO (significance level 2.2), SCIAMACHY limb (significance level 10.6) and SCIAMACHY solar OEM (significance level 6.6). Other data sets for which drifts are found compared to most other data sets are SCIAMACHY limb, SCIAMACHY solar Onion and SMR 489 GHz.