the Air Vent

Because the world needs another opinion

On line

Posted by Jeff Id on December 16, 2010

Our article on Antarctic temperatures has finally made it on line: Some results and ALL of the code is archived here:,  the SI can be down loaded from that link.

A detailed analysis is presented of a recently published Antarctic temperature reconstruction that combines satellite and ground information using a regularized expectation-maximization algorithm. Though the general reconstruction concept has merit, it is susceptible to spurious results for both temperature trends and patterns. The deficiencies include: (a) improper calibration of satellite data; (b) improper determination of spatial structure during infilling; and (c) suboptimal determination of regularization parameters, particularly with respect to satellite principal component retention. We propose two methods to resolve these issues. One utilizes temporal relationships between the satellite and ground data; the other combines ground data with only the spatial component of the satellite data. Both improved methods yield similar results that disagree with the previous method in several aspects. Rather than finding warming concentrated in West Antarctica, we find warming over the period of 1957–2006 to be concentrated in the Peninsula (≈0.35°C decade−1). We also show average trends for the continent, East Antarctica, and West Antarctica that are half or less than that found using the unimproved method. Notably, though we find warming in West Antarctica to be smaller in magnitude, we find that statistically significant warming extends at least as far as Marie Byrd Land. We also find differences in the seasonal patterns of temperature change, with winter and fall showing the largest differences and spring and summer showing negligible differences outside of the Peninsula.

24 Responses to “On line”

  1. John F. Pittman said


  2. andy said

    Well done

  3. Ryan O said


  4. lucia said


  5. Ryan O said

    Thanks, Lucia!

  6. Kenneth Fritsch said

    I plan on purchasing a copy of the paper and commenting on it here. It would be nice if there were a separate thread for this purpose. The publishing process I am sure has limited the scope of the paper, but there are other collateral issues such as estimating uncertainties as mentioned by Ryan O in another thread, that would make interesting discussion. Also it would be of interest to me to discuss how one would (if not in the paper) estimate the change in the spatial correlations over time.

    One thing I hope is made clear is that a paper of this nature must necessarily limit its criticism to a rather narrow band of the total package of issues that might be available for criticism or “improvement”. I assume that the RO10 authors had to assume that the satellite data was reasonably valid for use in correlations and did not attempt in this paper to evaluate how valid it might be. I think Nic L alluded to that data being unstable over time but stable over space at a given time – although I might be reading too much into his brief comment in a post.

    Thanks again to the authors for all the time spent discussing the issues at this blog and revealing the behind the scenes machinations of the publishing process. I hope the post publication process can spawn further interesting conversations.

  7. Carl Gullans said

    I asked this in another thread, and I think it was overlooked by more interesting commentary, but I’ll try again… how manual is the process that this paper uses to get its results? In five years, we will have more station data and more satellite data; is it somewhere close to trivial to get an updated look at things at that time, or will be there be significant burdensome manual intervention?

    Congrats on getting published.

  8. Jeff Id said

    The process is fairly automated except for the Satellite data which has been cleaned by unknown processes. Comiso has multiple publications on the matter, I just don’t know what was done to this data. The ‘cleaning’ is basically identification and removal of cloud contamination in the data. You could run the same process using the S09 satellite data and up to date temp stations fairly easily. The code is on line, basically turnkey, and linked above if you wish to try.

  9. boballab said


    Jeff, Avhrr instruments on satellites in the past have had random errors crop up due to their scan motors. The effect is that you get bleed over from other channels, so that might be one of the things that gets “cleaned up”. Classic example of this is the NOAA 16 Avhrr that NSDIC put warning out on:

    Note: Due to a problem with the NOAA-16 scan motor, all the channel data is shifted sporadically between 2001 and 2005, causing the channels to contain data from another channel; thus, the derived parameters also contain errors during this time-period. See Table 12 in the Error Sources section of this document for specific dates.

    This Scan Motor problem crops up on just about every satellite:
    NOAA 11

    09/13/1994 00:00:00 SCAN MOTOR Motor current increased and black body channels no longer operating within normal limits

    NOAA 12

    10/20/1991 00:00:00 SCAN MOTOR AVHRR sees temporary periods of high sync jitter and motor current changes due to possible bearing problems. Disabled TCE to increase lubricant.

    NOAA 14

    09/29/2005 16:00:00 SCAN MOTOR N14 AVHRR downgraded to RED status because of elevated scan motor current and persistent bar coded imagery.

    NOAA 15

    04/23/2001 00:00:00 SCAN MOTOR AVHRR data quality degrading
    04/30/2000 00:00:00 SCAN MOTOR AVHR scan motor current surges

    NOAA 16

    10/06/2005 03:00:00 SCAN MOTOR AVHRR scan motor current spiking and causing intermittent imagery bar codes again. Imagery clears up with each MIRP Rephase enable operation, executed once per orbit over the south pole but frequently returns a few minutes later. This is a chronic problem with this payload.

    NOAA 17

    03/31/2010 00:00:00 SCAN MOTOR AVHRR Product Data is now considered unusable due to increased degradation of the Scan Motor’s performance.

    Right now only NOAA 18, NOAA 19, and METOP-A are listed as having completely green AVHRR instrument packages. I don’t know if NASA’s AQUA satellite has that instrument package and if so what it’s status is, but as you can see from those quotes and if you click and scan the histories there has been lots of times they had to “clean up” the datasets from them.

  10. curious said

    Congrats to all – like Kenneth I’ll get a copy soon and get stuck in. Just hope I can work out which way up to hold the paper…

    I saw on another thread Jeff you are thinking of winding down the blog and you’d been hanging on for publication. I can understand that – after following along for a couple of years now, I sort of think that if there is really anyone out there who believes the “science is settled” then more blogging isn’t going to change things. Looking at the comments Judith Curry put in to the USHoR I think there is at least a reasonable chance things are going to change. FWIW, with the right people, I think that things could be knocked into shape pretty quickly.

    If you do carry on tAV I reckon there is a niche for putting up good quality posts on non consensus science in general. With the readership you have here it doesn’t take long for nonsense to get filleted and some good stuff could emerge.

    Thanks for all the education and fun! :-)


    9 Boballab – I think there was something on this a while back. I think it was at CA – I certainly remember someone posting a graphic showing the discontinuity in the records moving from satellite to satellite. IIRR Ryan picked up on it and I sort of have it mentally ticked off as having been bottomed out but I can’t remember for sure – No doubt someone else will!

  11. pete m said

    Congratualtions to O’Donnell et al. Wonderful to see your efforts rewarded. Please keep the blog going as your comments policy is refreshing amongst climate blogs and I’m selfish and don’t want to have to find another place where words are allowed to be published in this manner. And you’ve got the right attitude!

    Now waiting for MSM coverage in 3, 2, 1 …


  12. Layman Lurker said

    #9 & #10

    Yes “Jeff C” (not Condon) did a post on the sat discontinuities and Ryan modelled and corrected for the discontinuities is a later post.

  13. Geoff Sherrington said

    So do you get the coloured cover with Ryan’s tiles? That would be fair. Congrats to all.

  14. Brian H said

    ‘Grats, but based on behavior to date, I don’t think just a few publications like this, showing how science should be done and documented, will suffice to shame the miscreants who have been violently abusing the process into cleaning up their acts.

    Expect a sh**-storm of disparagement.

  15. Kenneth Fritsch said

    I am almost through my first reading of the “paper”, and while I’ll need another read or two to understand better the linear algebra and algorithms used, at this point, I am most impressed with the massive amount of work that must have been required to do this analysis and provide alternative methods. Reading the paper gives me a better (maybe more confident) reading on what Steig et al did and where they went wrong/astray/less-than-optimal. The paper in my view covers well the sensitivity testing of methods that I do not always see in much of the climate science literature. The methodological approaches used by the authors seem like common sense to me after, of course, the dissection of what Steig did.

    Using the satellite data for spatial reference, and not temporal, makes sense to me, as in the initial read I anticipated that was what Steig was doing. In the end, I am most puzzled why Steig did not do it that way. My puzzlement might well be a function of my not appreciating sufficiently the thinking that is required in applying these algorithms. Or could it be that Steig et al saw no (major) problems with the temporal part of the satellite data?

    I certainly hope that we see more discussion from TAV participants on this paper when it is finally read. The problem with blogs for slow thinkers like me is that by the time the light goes on in my thought processes, the blog has moved on.

  16. Kenneth Fritsch said

    something that cannot even be said of Nobel-caliber physicists on the skeptic side.

    You make a good point here and one I made to my brother, the lawyer, who read what the physicist (I think you have in mind) had to say (and his critics). My point to him was much the same as I am making to you and that is that I have heard brilliant scientists outside the field (climate science) making comments that completely miss the more nuanced situation which I judge exists. That happens on all sides of the climate science issues, but I think with the general consensus that exists when a skeptic scientist steps into it we are more likely to hear about it. I strongly suspect that specialists within the climate science community have some of the same problems with other specialists areas in climate science and just as a brilliant physicist can miss the point so can a brilliant climate science specialist.

    Judith Curry eluded to a similar point in that she more or less confessed to accepting what other scientists had to say about AGW because, well, they were scientists. How many times do we hear that a gaggle of Nobel, or at least prominent, scientists have agreed on an issue that is hardly in their field or specialty, but all the same we are expected to view that as something near conclusive on the matter. Most times it says more about their political leanings than any great insights they may have.

    The incentive of the scholar it would appear is to push the boundaries in his specialized field and he can easily get “lost” in that effort with regards to obtaining more generalized knowledge and even having a world view. That is not to say that all scholars or even specialist scholars are lost and, in fact, a goodly number are no doubt in the mold of the Renaissance Man/Woman.

  17. Kenneth Fritsch said

    The above post was a lazy one for the lazy thread. Sorry about that. I hope to have some questions about the paper soon.

  18. Kenneth Fritsch said

    I have read R10 and couple of times and have noted that with this paper’s comprehensive analyses of S09 and suggested algorithms to correct S09 methods that each read brings into better view some the points of contention. It is perhaps a bit too easy for a layperson, such as myself, to miss some of these more nuanced points and thus generalize too much in what I take away from this paper. Anyway, I would like to make a few points and ask a few question and do it in small chunks and not all together.

    I am assuming here that the critical aspects of your analyses and corrections deal with calibration, spatial structure considerations and determining the truncation parameter for PCA. I also assume that the latter 2 parts account for most of the differences between the results obtained by S09 and RO10.

    First, I would like to know how much computer power is required if I wanted to duplicate what you did in the algorithm calculations. I would want to attempt this, not to find any errors, but to get a better feel of how these algorithms are applied and learn in the process.
    Secondly, I would ask what parts of the methodologies that were put forth in this paper do you consider unique/original. I was not aware of the non-hueristic approach you showed with cross-validations for selection of truncation parameters being applied to PCA – even though you reference a paper by Mann suggesting that approach.

    My third part query will be about the assumption made for the Antarctica ground stations (GS) temperature measurements and the AVHRR (SAT) satellite data.

    It appears that both S09 and RO10 assume that the SAT data is stable spatially, but I am not certain whether S09 agrees with RO10 (and all the evidence provided in that paper) that SAT has shortcomings with temporal stability. A secondary question on this matter is whether it can be shown that SAT is reasonably stable in space over time, i.e. can it be assumed that the spatial relationships do not change significantly over time.

    RO10 uses the GS as a temporal reference for the reconstruction, but I do not recall if S09 was explicit about the necessity of this reference or did they equate the temporal accuracy of SAT with that of GS or even favor SAT over GS in this aspect? Also while there is good evidence that GS has better temporal reliability than SAT, is there reason to believe that SAT would be significantly less reliable than say temperature measurements from most of the rest of the world with respect to temporal stability?

    My final comment is not directed to the RO authors, but instead thrown out for general consumption. I am curious whether we might soon see attempts to do re-analyses of the SAT and/or GS measurements in the Antarctica. Or will we perhaps see an effort to discredit the GS measurements such that the error limits for a reconstruction, such as that performed in RO10 will be given much wider CIs and keep in play just about any conclusion one might want to make about warming (cooling) in that part of the world.

  19. Steve Fitzpatrick said


    I would bet that there will be some kind of a formal reply from the S09 authors (and perhaps others) to discredit RO(10). Of course, if it is a comment on RO(10), then RO et all will have a chance to reply…. so my guess is that they will attempt to publish a paper in another journal or a “stand-alone” paper in the same journal that tries to refute the results of RO(10)… we have seen this tactic before.

    BTW, my impression is that most of the difference in the reconstructions results from a difference in the number of retained PC’s.

  20. Ryan O said


    I appreciate all of the thoughtful comments, and I apologize I haven’t had time to respond to them up until now. I have a few moments, and would like to go through some things.

    First, I’m glad that you see the approach as more common sense and less elegance. Our basic thought behind all of this is that we know that there are errors in the data, but we do not have the expertise (or space in the article) to deal with all of them. Therefore, the approach was to maximize use of the data which we have the most confidence in (the raw ground station data) and minimize the impact of errors in the data for which the errors are likely to be larger (the satellite data). While it took quite some time for us to develop an approach that we all agreed on, that was the basic philosophy.

    In answer to your three-part question:

    1. Computing power: It helps to have 4GB of memory, but it is not necessary. The entire algorithm can be run on a single-processor Pentium with at least 1.5GB of memory. Some of the calculations for the responses I simply ran on a laptop at work.

    2. I’m not entirely sure what is “original”. I think the combination of methods we used in original, but individually, the various parts of the overall method have been used for similar purposes in the past. Perhaps the only original aspect is the cross-validation, as we obtained holdout statistics for every single station and every single observation for those stations. I do not believe any other reconstructions conducted as thorough of a cross-validation analysis as we did.

    3. The difficulties with using the AVHRR temporal data directly are fairly well-documented in the literature. Unfortunately, there are few (if any) means of calculating corrections to make the data directly usable. Trischenko has done a lot of work in this regard, and the sources of error on the AVHRR instruments are quite varied, and far larger than most people would assume. Fortunately, however, these errors will tend to affect the entire grid simultaneously (such as splicing errors), which would lead one to believe that the effect on the spatial relationships would be less than the effect on the temporal relationships. This, however, is not proven anywhere. We simply made this assumption.

    For reasons that I will not explicitly go into here, I believe that S09 felt the satellite observations were of comparable temporal accuracy with the ground station observations, at least in terms of trend. I do not know why they assumed this.

    I’m not entirely sure what you mean by “temporal stability” in this context. There are far larger opportunities for the satellite data to show spurious trends than ground stations, simply because the satellites are a single instrument and the ground stations are multiple instruments. Any error on the satellite (either due to instrument issues or data processing issues) are guaranteed to affect all measurements, and, depending on the type of error, may not affect all measurements in exactly the same way. In the case of ground instrumentation, however, no instrument is dependent on the error in any other instrument, so the impact of instrument error is far less. While there still may be systematic biases, the chance of those biases being of a significant magnitude are greatly reduced compared to the satellite data.

    The final comment is related to whether the spatial relationships are constant in time. This is not a terribly easy question to answer precisely, but it is not too hard to answer qualitatively. To test whether our spatial assumption was true (at least, true enough) we performed a number of tests. Some of these do not appear in the final paper:

    1. Use only portions of the satellite data. Even when restricting the satellite data to a period of 6 years (the smallest we went), the reconstructions do not significantly change. While this does not say anything directly about the pre-satellite era, it does indicate that the spatial relationships were reasonably constant during the 1982 – 2006 period. We had no physical reason to suspect that they would be substantially different in the earlier half of the reconstruction.

    2. Perform holdout calculations for pre- and post-1982 ground data. If the spatial relationships in the 1957 – 1982 period were substantially different than the latter half, we would expect significantly worse verification statistics when holdout testing was performed on the 1957 – 1982 ground data. However, we found that the verification statistics were quite similar in the two periods. So if the spatial relationships did change, they did not change significantly enough to undermine our conclusions.

  21. Ryan O said

    Steve, Ken . . . though you might think that the number of retained eigenvectors was the biggest reason (and so did one of the reviewers), it actually only contributes about 33% of the difference in regional trends. This is up here:

    if you want to see the sensitivity testing that we used to demonstrate that the reviewer’s supposition was not correct.

  22. Kenneth Fritsch said

    Ryan O, thanks much for the replies and particularly for sharing your sensitivity testing of the stability over time of the spatial correlations of the AVHRR data.

    A bit off topic but relevant to RO10 and some of what I have been attempting to muddle through is the comparisons of spatial correlations of temperature series between proxy and GHCN station pair-wise distance separations. In the Lungqvist thread LL gave a link to a post that you made here at TAV last March. I re-read that post and it indeed does provide a means of putting that correlation versus distance into a global scale model by using the more complete UAH data. In fact, I have seen other publications for doing what you did (with the intent of determining the uncertainties of surface instrumental temperature measurements – given the spatial and temporal infilling required) that is not as complete as what you did in that analysis. I have not found any authors who have looked at the satellite data as you did, even though the completeness of coverage would seem to make that data a prime target.

    Anyway I find that there are significant differences in the above relationship between GHCN station and proxy data even when ignoring the factors of latitude that you include in your equation. I think my next step should be a comparison of UAH (or RSS), GHCN station and Proxy data with the series correlation versus distance separation and then looking at these three sources of data in light of ARIMA models. I may want to “borrow” your equation for doing part of that analysis.

    Again, I think that the comprehensive analyses and explanations of methods that appear in RO10 could be a benchmark for other papers coming out of the climate area of science. I plan to ask some more questions in the future about RO10.

    Also, I have noted that I should be talking about the Antarctica continent, but it should be Antarctic PCA or Antarctic winter.

  23. Ruhroh said

    Hey Jeff;

    Did you all remember to separately homogenize the Tmin and Tmax in the approved fashion?


  24. [...] “peer review” and published in a climate journal.  Eventually they did so – publishing a rebuttal just before Christmas, but yesterday it emerged that Steig was a reviewer of the paper [...]

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s


Get every new post delivered to your Inbox.

Join 133 other followers

%d bloggers like this: