A first look at UAH 5.3

Readers here (and elsewhere) have noted the release of a new version of the UAH satellite-derived tropospheric temperature record by John Christy and Roy Spencer. As noted by Spencer on March 5, version 5.3 seeks to remove, or at least mitigate, a spurious annual cycle that was apparently exacerbated in the changeover from the older MSU sensors to the newer AMSUs back in 1998.

The new data set is now available, so I’ll show some graphs that highlight the differences between the two versions. But first I’ll give a brief background on the matter, including my role.

[Update, March 10: It appears that John Christy was first notified of the annual cycle issue in October 2008, although it is unclear whether he understood the implications at that time. ]

A possibly spurious annual cycle in the UAH data set was noted by Eric Swanson as far back as 2003. I first became interested in the issue after reading detailed descriptions of the UAH annual cycle by Tamino in Open Mind, back in October 2008. (The blogger Atmoz had also posted on this previously, as it turned out).

[Update, March 10: Tamino’s first post came on October 21, 2008 on the matter was entitled simply UAH and RSS. In that post, Tamino plotted the difference between the UAH and RSS monthly record and noted two anomalies, namely a “step” around 1992, and a strong annual cycle in the UAH-RSS difference over the last decade.

A Fourrier frequency analysis of the two series for 2003-2008 showed similar spectra, except with the addition of a very strong annual cycle in UAH (the blue peak at one year).

Tamino spent most of the post analyzing this annual cycle, concluding:

So it’s the UAH data that show a false annual cycle recently, in fact they show a semi-amplitude 0.13 deg.C, full amplitude 0.26 deg.C. It’s just not believable that the annual cycle in global lower-troposphere temperature has changed its amplitude by 0.26 deg.C between the 1979-1999 reference period and the 2003-present period. Something is wrong with the UAH reduction.

In comments, Richard Steckis said he had run the post past John Christy and that Christy had responded that the “step” difference was a problem in RSS, not UAH, citing a number of references.

The evidence is pretty clear that RSS has a spurious warming shift in the 1990s. Some of that information is in the three papers attached. The key point is that RSS shows a jump relative to all other datasets (UAH, SSTs, HadAT, RATPAC, surface temps, US radiosondes, Australia radiosondes etc.).

Understandably dumbfounded by Christy’s failure to address the clearly spurious annual cycle, Tamino responded:

Response: Did he have nothing at all to say about the spurious annual cycle in the UAH anomalies? …

Later in the thread, Robert Randall (co-author of a JGR paper on UAH/RSS differences) appeared, agreeing with intervening comments from Steckis that it would be “premature” to ascribe the problem to UAH:

The annual cycle signature can be explained. The diurnal and the hot target corrections that are applied to the raw MSU data have an annual cycle in them. Therefore, as you stated, the differences in the databases will have an annual cycle in them. As the process for determining the diurnal cycle correction is different for both groups there is a temporal signature that is caused by this difference. It might be premature, however, to claim that it is not possible for the amplitude to increase to what you are showing, thus concluding that UAH has the problem. It is possible that the actual diurnal/hot target correction required is at that magnitude.

So it does appear that Christy was advised of the issue at the time, although it is unclear if he understood the full implications of the problem. In any event, there were no further comments from either Steckis or Randall, even after Tamino’s detailed follow up post, which clearly demonstrated that the UAH annual cycle was spurious, nonsensically appearing over ocean, in the tropics and in the southern hemispehere. And there the matter lay for several months.]

I subsequently wrote about this three times in all, including this key post where I showed how the UAH annual cycle (which peaked in the months of January and February, and troughed in May or June) resulted in trends that vary strongly by month and season. I showed this variation, which has no possible physical basis, in the following graph.

Each temperature data set was subdivided into twelve time series, one for each month of the year, and then the trend slope was calculated for each month. Note that UAH monthly trends vary much more than those of the other data sets, including  RSS which is also a tropospheric temperature record based on the same satellite source.

In July of last year, Anthony  Watts discussed a high value in the NASA GISTemp June 2009 surface temperature anomaly, which struck Watts as unrealistic, especially compared to the UAH value for that same month. One reader, Paul K, posted a link to my blog, and opined that at least part of the difference could be due to  the regular “dampening” at mid-year in UAH.

Every year the UAH data show a substantial drop in May and June. There is a serious seasonal variability in the UAH data, and it seems to be getting worse. For some reason UAH shows a seasonal rise in the anomaly in February, and seasonal decline in the anomaly for May and June. This has been discussed at several blog sites, such as

I added a comment later on, pointing to my previous more substantive posts, as well as noting Eric Swanson’s and Tamino’s earlier references to the matter.

Watts soon got in touch with John Christy, who acknowledged in a July 2009 “readme” file that the annual cycle problem should be addressed, as well as elucidating its source.

It was brought to my attention by Anthony Watts that there has been some discussion about the noticeable annual cycle in the LT and MT trends when done by months. In other words, the trend for Februaries is on the order of 0.12 C/decade warmer than the trend for Mays.

The feature arises when the AMSU data are adjusted and merged into the MSU data stream beginning with NOAA-15 in Aug 1998, then carries forward with NOAA-16 and AQUA (both of which are AMSUs too). The process involves at one point the removal of a mean annual cycle in the anomaly differences from one satellite to another. It turns out that all satellites have a residual annual cycle due to each instrument’s peculiarities. In the end, all annual cycles are matched to NOAA-6 and NOAA-7 …

I’ve tested a number of alternate processing methods (basically versions of not removing the annual cycle in the difference time series from the first AMSU onward) and the range from the highest to lowest is reduced to just under 0.09 C/decade. This in effect establishes a new annual cycle for the AMSUs based on the first AMSU.

I think the magnitude of the annual cycle in the monthly trends is alegitimate problem to address. The range in the current v5.2 LT looks too large (about 0.12 C/decade) …

That proposed fix has now been implemented. The new data set is virtually unchanged from the old one up to mid-1998; so without further ado, here is a comparison of the two versions from 1998 on.

One can readily see that version 5.3 is somewhat lower at the beginning of each year, and higher at mid-year, just as one would expect. Here is the difference between the two versions, showing the annual cycle that has now been removed.

Finally, I have recalculated the trend in global lower troposphere temperature for each month, and compared the two UAH versions to an updated RSS plot. In general RSS trends are higher, of course, and there still seems to be somewhat more month-to-month variation in the UAH trends.

The UAH peak trend is now in the months of September through November, while February is considerably lower than it was (0.15C per decade, down from 0.2C). June now has the lowest trend, with May bumped up from just over 0.06C to almost 0.09C per decade). Overall, the peak to trough difference in the UAH monthly trend is at 0.08C in version 5.3, down from 0.14C differential seen in the previous version. RSS peak-to-trough trend differential is at about 0.06C. So while UAH still has more seasonal variation, the contrast is much less severe than it was before.

As expected, the overall global trend for UAH remains the same at 0.13C per decade since 1979, a little lower than the other temperature data sets. In a future post, I’ll update the monthly trends of the surface sets, as well as look at the tropical zone and other aspects of interest.

Finally, I should briefly comment on the suggestion that I have not received proper credit for my role in identifying the annual cycle issue.

As I mentioned above, others had identified this issue first. And, although I was intending to contact John Christy about it, I did not do so until after Anthony Watts did. So, even though Watts in no way discovered or characterized the annual cycle, he may well have been the first to notify John Christy about the problem. [Or not: See above.] Of course, it’s problematic to acknowledge anonymous bloggers in any case.

It’s interesting to note that, aside from Swanson’s 2003 paper on Antarctica trends, the annual cycle discussion occurred mainly in anonymous blogs. I would say my contribution, as such, was to show how an ever-growing annual cycle led to large and undoubtedly spurious discrepancies in seasonal trends. It may well be that this practical demonstration was convincing in some way.


14 responses to “A first look at UAH 5.3

  1. This still doesn’t seem right to me. Why are they adjusting the Jan / Feb temperatures? Looking at your first graph, these seemed to be correct all along.

    They appear to have made these adjustments to keep the somewhat lower trend of 0.13C/decade as is, but isn’t the reason that the trend for UAH is lower than the other datasets because of their strong seasonal cycle in the first place?

    [DC: I’m not sure why UAH is lower than the others, but the particular adjustment that has now been undone would not affect the overall trend. Also note that there are probably multiple components to the seasonal cycle, some spurious and some “real”. ]

  2. carrot eater

    The seasonal variability is reduced; can we get a metric of how much? Maybe the standard deviation of the monthly trends, compared to RSS?

    The Watts post was just weird because it was comparing anomaly magnitudes straight-up, without correcting for the difference in baselines between UAH and GISS. A recurring issue over there, it seems.

    [DC: Yep. The classic is the anomaly histogram “analysis” that Watts did back in 2008. The actual difference between GISS and UAH in June 2009 was reduced considerably, once you use the same baseline, and remove the UAH annual cycle. ]

    • Concerning your first point, it would be interesting to compare the standard error of the monthly trends to a measure of the spread of those trends. Or else pairwise correlations between the monthly series within each data set. I’m open to suggestion on that.

    • Using the simple regression model (which is probably not far off since there are only 30 data points), these monthly trends have 95% confidence interval of +-0.07 or 0.08C per decade. So when the spread gets up to 0.14C, as in UAH 5.2, the peak and trough trends are well outside each other’s confidence intervals.

  3. It’s also worth noting that Christy didn’t actually credit Watts for finding it (nor suggest that Watts was trying to take credit for it). . . just for notifying him that there had been “some discussions” on the web about it.

  4. DC, thanks for clarifying your role in this, as well as shedding some light on the history of the UAH data.

  5. >”Also note that there are probably multiple components to the seasonal cycle, some spurious and some “real”.”

    I would assume that at the start of a GHG induced warming you would expect the highest trend in Sep-Nov. Presumably this would lessen as the GHG induced warming became more mature?

    [DC: I think warming was expected to be greatest in NH Winter (i.e. Dec-Feb.). However the UAH cycle was extreme and even showed up in zones where it didn’t make sense (i.e. tropics and SH).

    Recently, Sep-Oct-Nov seems to have come to the fore, as I pointed out a while back in regard to NASA GISTemp (see first chart above). This appears related to rapid warming in the Arctic in those months. See this post by Tamino which was largely inspired by my previous comments. ]

  6. Talking about warming: I’m a little bit amazed at how the 2010 anomaly trend at the daily UAH global temperature trend website keeps cruising in 20 year record territory. I wonder when it will go down again.

    No active sun, a negative PDO (correct me if I’m wrong) and an El Niño that is considerably weaker than the Super El Niño of 1998, and still (near-)record warmth, globally speaking. How can that be? 😉

    [DC: And given the six-month lag in El Nino effect on global temperature, we still haven’t seen the peak, even if it is fairly weak. Yet it does look as if 2010 will be #2 or even #1 in some data sets.

    I couldn’t repair the link because I didn’t see the URL target. ]

  7. There is a major update above, summarized in the opening:

    Update, March 10: It appears that John Christy was first notified of the annual cycle issue in October 2008, although it is unclear whether he understood the implications at that time.

  8. >”[DC: And given the six-month lag in El Nino effect on global temperature, we still haven’t seen the peak, even if it is fairly weak. Yet it does look as if 2010 will be #2 or even #1 in some data sets”

    I could easily be wrong but I understood there is a 6 month lag of global temperatures behind SOI, but the lag is much smaller for other indicies like ONI/MEI. SOI had a minimum in Oct but is now more negative. ONI hit its peak at NDJ and the models have a strong consensus on declining values. MEI is still increasing and there are hints of the lag only being 0.5 months. What lag to which index should one choose?

    [DC: The lag looks to be about for 3-4 months from ONI peak to RSS peak for strong El Nino (1998, 2003). So I stand corrected – I would expect FMA or MAM to be the peak rolling three-month period. Didn’t check the surface record though. ]

    Perhaps the link was

    [DC: That’s the one – be sure to use ch 05.]

  9. Oh and while mentioning ‘ 2010 anomaly trend at the daily UAH global temperature trend website ‘ and the 20 year record highs

    What does 20 year record high mean? I see that parts of 2004, 2007 and 2009 and other years are higher than 20 year record high at some times. Does this imply 20 year record highs is of 1979 to 1999? Hmm no there are brief times when 1998 is higher. So what does it mean?

    [DC: 1998 does seem to follow 20 yr record high very closely from Aug. to Nov. then drops off. So maybe it is 20 yr baseline, 1979-1998. ]

  10. Aww, wasn’t that a product of continental distribution?

  11. The effect of the new version would seem to prolong the meme that temps haven’t risen since 1998. In the new version 1998 temps remain about the same, while 2010 temps have been downgraded.

    Previous version:


    Jan – 0.58
    Feb – 0.75


    Jan – 0.72
    Feb – 0.74

    Updated version:


    Jan – 0.58
    Feb – 0.76


    Jan – 0.64
    Feb – 0.61

    (the old version only goes up to Feb 2010)

    Don’t know if there’s much difference in the trend for that period, but if 2010 ends up being lower than 1998, ruler-wielding denialists will no doubt make use of that.

    [DC: I’m pretty sure that 2010 will end up #1 or #2 in most data sets, unless ENSO flips into La Nina really fast.

    The adjustments in UAH 5.3 should not affect the annual average at all. The Jan-Feb peak is now lower, but May-June will be higher than it would have been (you can see this in 2009). ]

  12. DC

    Looks like the correction has moved all the way through to the Graph that shows daily AMSU ch5 readings… Would be interesting to see what happened as most of the ‘hottest days recorded’ this year are no longer the hottest…

    [DC: I’m not sure what is going on there. It could be that monthly adjustments are applied retroactively to the previous month, so that the data is no longer “raw”. Certainly, a good question for Roy Spencer. ]