the Air Vent

Because the world needs another opinion

Lucky Number 0.07

Posted by Jeff Id on October 11, 2009

We’re getting closer to publication on the Antarctic trends. There are still a few details being worked out so this is not the final result, however thousands of people read here so it’s important to provide an update of the results and a comparison of the similarities and difference from Steig 08.

From the abstract of Steig et al.

Here we show that significant warming extends well beyond the Antarctic Peninsula to cover most of West Antarctica, an area of warming much larger than previously reported. West Antarctic warming exceeds 0.1 6C per decade over the past 50 years, and is strongest in winter and spring.


West Antarctica warmed between 1957 and 2006 at a rate of 0.1760 +/-.06 C per decade (95% confidence interval). Thus, the area of warming is much larger than the region of the Antarctic Peninsula. The peninsula warming averages 0.1160 C per decade. We also find significant warming in East Antarctica at 0.1060 C per decade (1957–2006).

Confidence intervals were edited out.

The main revelation in the paper was reported to be an increased warming of the West Antarctic as compared to other results. However, most of us skeptics are more interested in the continental trend. From the area weighted reconstruction which is based on surface stations within the Antarctic continent the trend we find is 0.064 C/Decade. The trend distribution looks like this:

Figure 1 – Area weighted reconstruciton using knitted (offset) surface station information to splice short stations into the complete record.

The method is simple and is a target pattern for the trend magnitude and distribution. The deep red of the peninsula section at the left is colored in the + 0.3 to 0.7 C/Decade range and is populated by over a dozen surface stations. In Steig et al only 3 pc’s were used to describe the continent, the low pc count necessarily results in the blending of surface station information around the continent. It doesn’t necessarily result in an incorrect average trend however. In this case though, there is a high concentration of surface stations in the peninsula region who’s information is blended across the continenet. I believe I’ve shown that reasonably well in a few posts including this post here. I need to mention that NicL did an excellent post which contradicts that conclusion. While I found Nic’s results accurate, I don’t interpret his method the same way. His post is here.

The point of this rehashing of the discussion occurs because a long time ago, back in February JeffC provided a regridded form of data based on hexagon gridcells with the purpose of reducing the effects of oversampling a single region of the antarctic. Since these results used the PC’s of Steig et al right from his published AVHRR data and Matlab RegEM, it represents a good comparison to the original results and no changes in understanding since that time have had an effect on the result. Please excuse the crappy old style graphic, note the trend in the figure is 0.07 Deg C/Decade:

Figure 2 – Trend 0.07C/Decade

Interesting that we achieved such a close match to the offset surface station trend with only 3 pc’s by simply making sure the peninsula information wasn’t recopied over and over prior to regression.

Below is another reconstruction by Jeff C, also back in February. This one used manned stations to infill surface stations. Steig et al reported a trend of 0.138 C/Decade for the AWS reconstruction but again simple regridding to insure that high density areas didn’t overweight their import in the reconstruction resulted in a trend of approximately 0.07 C/Decade.

Figure 3 – AWS Reconstrucitons from Surface Stations.

It’s clear from these examples that the trends are sensitive to the density of stations in low PC RegEM. This is pretty obvious conceptually at this point, but may be confusing if you’re not familiar with our work here. What’s more is that I believe it’s strong evidence supporting my station weighting conclusions that the peninsula trends are driving the trend over Nic’s version.

Below is the latest emulation of Steig et al by Ryan. Ryan figured out that in using correlation scaled satellite data the resolved PC’s are a match to Steig and replication is now near (but not quite) perfect. Many of us have seen the Antarctic temperatures from Steig et al and their bright red colors, however when the results are plotted on the same scale used in the surface stations above the story changes a bit.

1957-2007 Steig replication 0.7 scale
Figure 4 – Steig et al emulation

Note that the trend is actually higher than Fig 1, but the blending of the trend across the continent is pretty clear. It’s also clear that the Peninsula warming is heavily understated compared to the oversampled surface station information in the peninsula. If you haven’t followed the whole thing this next graph might not make sense bit it is a plot of the fitted data only rather than a combined plot of AVHRR and fit data. The S08 results used raw satellite data from 1982 onward instead of displaying the fit results. In the end this resulted in a further distortion of trends as the satellite data is unbelievably noisy due to cloud contamination and had an anomalously high trend in comparison to the surface station data. I think it was done because RegEM doesn’t make it easy to get the full timeframe of the reconstruciton but it is one of the problems with S08.

1957-2007 Steig replication no AVHRR 0.7 scale
Figure 5 – S08 emulation – no direct AVHRR data.

Note the change mutes the trends considerably, however the pre-satellite data portion of the reconstruction dominates the trend.

We’re still discussing the final version of the improved reconstruction for publication, however we’re down to about a tenth of a degree difference between any method we choose. Two of the several methods are presented here, one was an eigen weighted EM reconstruction which weights surface stations according to their eigenvalues. It’s basically an iteration improvement in the unweighted EM version which changes the final result slightly. Most of the difference occurs because of the inclusion of what we believe is the correct number of PC’s. You cannot include this many PC’s in the RegEM version without overfitting so this represents a significant improvement over the RegEM method. Also, these methods have basis in previously published literature, nobody’s inventing the (magic) wheel.

1957-2007 Eigenweighte EM 0.7 scale
Figure 6 – Eigenweighted reconstruction.

You can see that the high PC count has localized the peninsula data much better. This is important IMHO as my view is that this information drove the high trend of Steig et al, at least in the pre-1982 portion of the reconstruction.

In addition Ryan has employed an improved method using Regularized Lease Squares on the left singular vector (the spatial distribution ) from a principal components analysis to fit the surface station data. This isolates and improves the response to the effects of known large discrepancies in the satellite trend information created by several factors, including steps between satellites, cloud noise and satellite temperature drift.

1957-2007 infl=1_3 RLS
Figure 7 RLS Reconstruction

A table summarizing the various trends by region is shown below. There is a lot of detail here.

Steig et al Emulation Regridded Steig Area Weighted Eigenweighted RLS
Continent 0.12 0.118 0.07 0.064 0.064 0.076
East Antarctic 0.106 0.097 0.025 0.042 0.025
West Antarctic 0.176 0.192 0.14 0.097 0.188
Ross ice shelf 0.219 0.24 0.009 0.127
Peninsula 0.116 0.126 0.358 0.298 0.427

Note that the AWS reconstruction trend of 0.069 was not included as it is not a satellite reconstruciton, however it matches the many methods we have tried and the four on the table right well. There is a large difference in trend between S08’s three PC version and the various other reconstructions. Over the past 9 months we have verified through several different methods that the actual continental trend of the Antarctic temperature is very close to 0.07 C/Decade from 1957 – 2007. People want a confidence interval for that so I’ll say it’s +/-0.1 ish so the trend is not significant. However, please don’t misunderstand the meaning of significance. The trend is real to a high degree of certainty, significance just determines the likelihood that the trend is caused by shorter term signals. The assumptions about significance in this case of normal distribution of weather and signal are probably not terribly accurate but it gives you an idea of the magnitude of high frequency signals on this data.

The 1982 onward trends are equally important however, and they have a slightly different look from Steig et al.

area wt 1982-2007
Figure 8 – Area Weighted reconstruction post 1982

1982-2007 infl=1_3 RLS
Figure 9 Regularized least squares reconstruction.
1982-2007 eigen em
Figure 10 Eigen weighted expectation maximization 28 PC

And of course, Steig et al.

Steig 09 1982-2007
Figure 10 Steig et al trend per decade

The point is that first Steig et al overstates the trend post 1982 by what is very likely a statistically significant amount in comparison to weather noise, however it is certainly a significant amount in comparison to the measurement accuracy. In my opinion it’s likely out of whack from the measurement accuracy by several sigma.

However despite the mathematical errors and trend blending, the warming of the west Antarctic region (considered the lower left) does indeed seem to be real.

27 Responses to “Lucky Number 0.07”

  1. Phil said

    Has Comiso released his cloud data yet? IIRC, Comiso had previously published a study that showed Antarctica to be cooling slightly, but in Steig, et al 2008, they added another filter throwing out any data that was +-10 deg C, so wouldn’t any reconstruction (yours, Ryan’s, Jeff’s or Steig’s) be sensitive to the new enhanced cloud masking?

  2. Jeff Id said

    It wouldn’t hurt to have better quality data but different masking for the most part only affects how the surface stations are weighted for area and shouldn’t have a huge affect on trend.

  3. curious said

    Hi Jeff – thanks for the update, looking forwards to the paper.

    “In my opinion it’s likely out of whack from the measurement accuracy by several sigma.”

    Are you going to mention/expand on this in the paper? Also are you going to give any context on Antarctica from a historical perspective? I think TonyB mentioned he was working on something? Personally I find it hard to be alarmed by 0.07degC/decade (slowing to 0.046degC/decade), especially whilst there is still so much to know on all this climate stuff.

  4. Jeff Id said

    It’s up to others how this is included. The paper is already long and being truncated for linguistic extravagances like adjectives and punctuation. Trend’s as small as 0.046 C/Decade are definitely nothing to fear so it’s hard to imagine the doom and gloom from shrinking glaciers based on that. I mean if the antarctic is 20C below zero on average, how the heck can 0.05C have any meaning?

  5. rephelan said


    I’m really glad to see you’ve decided to move toward a publication. Good luck.

  6. curious said

    Agreed re: temps.

    How is it going with costs? Is Lucia still collecting?

  7. PaulM said

    Nice summary Jeff. Are you going to post up the paper here before submitting it? You could get some good instant peer review that might lead to a better submission. You’ve done so much stuff on this, it must be tricky to decide what to include and what to leave out. IMHO the most important thing is the effect of increasing the number of PCs and your sentence “the low pc count necessarily results in the blending of surface station information around the continent”.

  8. Jeff Id said

    Thanks guys, please understand this is almost entirely Ryan’s work. Choices he makes on what to disclose are going to be based on Journal policies which often include not being scooped by another publication. I’ve done nothing WRT costs yet but we will need some help. Currently it’s just about getting the final details completed and I just couldn’t wait to show people how amazingly consistent the results are.

    Certainly the last Steig plot is way out of whack. It’s too bad they didn’t replace the AVHRR with the reconstruction data but I agree with Paul, the low PC count is the problem. No conclusions about spatial or even continental trends can be taken from S09, but it is such a cool method they used that by improving on it using several different mehtods we actually get a decent and stable result. I was quite skeptical that could be accomplished, but through several careful changes to the algorithms Ryan (and Nic) have been able to mathematically narrow down the number of PC’s to 28 before overfitting occurs and have verified the result through a variety of methods.

    You can really get bogged down in the details of these things but stability of result to different methods is key. I’m thinking of a gavin quote about robustness in publication or something – but I’ll hold my tongue. Incidentally after this post, my positive matrix factorization failures may not have been as bad as I thought and deserve another crack too.

  9. treyg said

    Could someone provide a few links that describe the motivation for this work? I’m new here — a little history on the subject would help. Thanks

  10. Jeff Id said

    #9 Here’s a good place to start. The air vent is loaded with posts on this paper.

  11. William said

    What do you think the odds are that a journal will publish your paper? Based on comments on WUWT the peer review process seems to be entirely co-opted in favor of AGW and there would not seem to be any chance of a favorable review.

  12. Jeff Id said

    I’m not really sure. I’ve seen some reviews from McIntyre’s latest work and they are clearly biased. There is no sanity whatsoever in them. At the same time, the methods used in our version are more commonly used than I had realized and in a perfect world the field would welcome the improvements. I wonder what the Air Vent would say if the paper get’s an unreasonable reception.

    We still show some warming, which is very much in line with a lot of previous work. Also, we show the western (southwest really) Antarctic warming that Doc. Steig was concerned with. However, the plot on the cover of Nature is not correct and Antarctic temps continent wide are exaggerated. Since something like 85% of the worlds land ice is there, it’s quite difficult to make a good case for melting of Antarctic land ice – not that that will stop the grant hunters.

  13. curious said

    Jeff – is it worth going for publication in a maths/stats/discipline other than climate journal?

  14. Jeff Id said

    Personally, I’d like to see the response from a climate journal first but these things aren’t up to me.

  15. Jeff C. said

    Jeff – thanks for the link to this post from the other thread.

    I have been following the progress and am looking forward to the finished product. I did that first “quick and dirty” AWS gridded recon on a Saturday afternoon back in February. It has been amazing to see the same 0.07 number come up over and over using much more rigorous methods.

    Two other early observations that seem to have held up:

    1) Pre-1980 warming is greater than post-1980 warming in virtually every alternate reconstruction. This was strikingly apparent in the AWS recons as post-1980 indicated substantial cooling, particularly in the gridded recon. The AVHH alternate recons are flat or show diminished warming post-1980.

    2) Dr. Steig’s methodology results in more dramatic warming than every reasonable alternate reconstruction. This was seen in the 50 year linear trend (0.07 vs. 0.12 deg C/decade), but also in the shorter duration trend over time. Steig’s recon didn’t show diminished warming post-1980 but indicated a relentless upward trend (that is, the post-1980 trend is even more extreme than the pre-1980 trend). As far as I know, none of the alternate recons show anything like this. I still don’t understand why his methodology appears to smear the exaggerated warming into the later decades of the record. Any ideas?

  16. Jeff C. said

    Ooops AVHH in the last comment should be AVHRR.

  17. Jeff Id said

    #15, We’ve got these things all worked out now.

    RegEM is a regression method which linearly combines surface station data to match the satellite data. It is regularized in TTLS by truncated PC’s. In RegEM the official version, only missing data is added to the matrix so the output post 1982 is the AVHRR data only. As you know the AVHRR is useless for trend information because of the satellite offsets you pointed out first. Ryan later ran some Wilcoxen test to determine that the satellite was significantly different from surface station data and developed some methods to correct for it.

    Current methods do not use the correction for clarity and defensibility in the paper but I think the corrected version would produce the most accurate results. It turns out that RegEM creates all the data for the best fit of surface to satellite and the matrix is masked by the available values so we only see the missing values change. In our current reconstructions we use the entire expectation maximized matrix for the result and copy it over the satellite data. In this way our trends are not distorted (very minimally distorted) by the satellite trend and the covariance information (the important part) of the spatial distribution of the AVHRR information is used to allocate the surface stations around the continent. This method has precedent in other publications.

    In the end we get a flat or cooling trend post 1982 which is correct in comparison to the simple area weighted reconstructions. Incidentally 1957 turns out to be a bit of a cherry pick as well as there is a pretty big dip in temps in the beginning of the record which is largely responsible for the trend. I’m told that going back further can really reduce the net trend but there are less stations available. I should try that sometime.

    Also, Ryan did a reweighting of the surface stations prior to iteration by eigenvalue at that point in the particular gridcell. If there are 28 pc’s and each one has a weight at that surface station point it gives a start weight for the iteration. This helps the algorithm have an improved global minimum which shows up in the verification stats. Another improvement was to solve the ground stations first while maximizing verification stats before solving the satellite matrix. The satellite matrix is then solved 1 PC at a time to prevent cross contamination of the non-orthogonal (because it’s not ideal) regression. It’s a little confusing but the result is that we can now regress up to 28PC’s without overfitting or adding noise and the whole thing has precedent in publication.

  18. Jeff C. said

    #17 Thanks for the explanation. It doesn’t take much to fall behind in the discussion around here.

    I had forgotten that the post-1982 trends was driven entirely by the AVHRR data reduced to 3 PCs and that the surface station data is used only to extend the 3 PCs back to 1957. Need to dig out the flow chart to refresh my memory.

    So was there ever any clarification from Comiso or any of the others regarding the methods used for cleaning up the AVHRR data? We both spent a lot of time working with the AVHRR data from the University of Wisconsin and could never get anything that looked like Comiso’s data. The UW AVHRR data had ugly discontinuities at the satellite transitions (as you mention above). It also had issue with inconsistent cloud cover masks and temporal and spatial data gaps. I wrote Dr. Key at UW about the problems and he readily acknowledged them. He also added that they weren’t quite sure how to fix them.

    The Comiso AVHRR set was beautiful, no discontinuities, no gaps, and ever increasing temperatures. Not to mention that the 2008 Comiso dataset strongly contradicted what Comiso himself published in 1999. It also was at odds with Monaghan’s very detailed 2008 recon (Comiso was a co-author) which indicated post 1980 cooling everywhere except the peninsula and a portion of West Antarctica.

    I wrote Comiso back in April and never got a response. Did anyone ever find out how he turned that ugly AVHRR data into such a masterpiece?

    Regardless, it seems clear that you and Ryan have a better method. Perhaps publication will prod the team into an explanation rather than the usual brush-off.

  19. Geoff Sherrington said

    I still think that “Ryan’s Tiles” showing the results of subjective options on regpar and PC numbers should be submitted to “Nature” as a counter. But so it be.

    On first reading of Dr Steig, there was reference to new methods to differentiate cloud covered areas from others. Has a more detailed description of methodology emerged?

    Last, I can’t find much confirmation of warming in data from manned Australian Antarctic stations up to the present, especially in the last 40 years, but these are a small subset and might not be representative.

  20. Jeff Id said

    #19 I’m curious if there is some additional station data available that is not part of the current dataset. My opinion is anything which could be added in would be useful.

  21. Layman Lurker said

    Jeff, you might want to check this out. I emailed it to you a while back:

  22. Jeff Id said

    #21, I saw that paper when you sent it before, it’s conclusion presented in the abstract is correct as far as I can tell.

    #18, I’ve had no communications with Comiso, Ryan was able to get a link to data I believe but I need to ask again because it’s been too long. The dataset was too large for my laptop to process so I haven’t spent the time to work on it. I’ll ask Ryan again b/c I’m not entirely sure what happened.

  23. Geoff Sherrington said

    #20 Jeff Id

    The Australian Bureau of Meteorology with others runs stations on the edge of the land at Casey, Davis and Mawson (shown on Google Earth). Some started in 1957 or so with IGY. Some had small station shifts but the public metadata I have seen is not so complete. They also run Macquarie Island which is half way between New Zealand and the Antarctic. The data are given as daily Tmax and Tmin but they might have been homogenised lightly by the BOM. If you have this datset, then I can’t add to it, sorry. Also, copyright is strict on this dataset but a few people like GISS and KNMI might have made further adjustments. And CRU???? I’d go for the earliest BOM data for these stations.

    (I’ve been looking at these places, trying to track strong global patterns in the anomalous 1998 hot year to try to explain its mechanism, as this potentially leads to info about mixing, lags, mechanisms. Do you know of anyone else who has looked at 1998 on monthly or better resolution?)

    Regards Geoff.

  24. John F. Pittman said

    JeffID, I saw an interesting comment, can’t remember where, that by using diferent models and selecting models based on diminishing CI’s was incorrect, because you have to include the CI’s of the models you rejected. This is essentially for the same reason that preselecting tree thermometers is an incorrect procedure. I think there is a “truth” to this. However, that same truth would indicate that whether it is tree pre-selection, or you or Mann it would mean the same thing. It was strange that the person should fault your approach, in that you are in a matter of months, following Mann’s procedure that took years. And he starts with pre-selection. I wonder just how large that would make his CI’s for his last paper. Say, “”The Mann Continuum, Confidence Intervals for a Multi-Model Temperature Proxy Approach 1998 – 2009.”” 😉

    If such selection of models is known to expand CI’s to infinity, how come that is not known for tree thermometers from pre-selection? There appears to be some self selective blindness going on.

  25. Layman Lurker said


    When doing a presentation, one might need to borrow Al Gore’s lift truck to point out the upper confidence limit. 🙂

  26. John F. Pittman said

    Layman Lurker apparently this is what NAS and Wegman said. 😉

  27. Ryan O said

    Jeff, I’ve got the row-centered version of RLS working. Much less sensitive to the period of satellite data you’re using than anything we’ve done so far. You get pretty much the same result whether you use the whole time, half the time (1982-1994 / 1994-2006), or single satellites. This lends credence to the idea that the high AVHRR trend is an overall satellite splicing problem rather than a cloud masking problem (which would be regional and would not be improved by global row centering). Blah blah blah . . .

    AFA the questions about Comiso go . . . I gave up trying to get anything out of him. He either did not respond at all (several times) or gave me a brusque who-are-you-and-why-should-I-waste-my-time-talking-to-you response (twice). Steig did give me links to get the raw AVHRR data – which I obtained – but the whole set is about a terabyte and needs lots of processing that I might never get to.

    However, I don’t know how valuable that would be . . . since I don’t really think the problem is with cloud masking. I think it’s with how the satellites were stitched together. Guys like Trischenko and Li (can’t remember which post I provided links to their papers in) feel the same way.

    Regardless, it’s pretty easy to show the divergence between ground and satellite temperatures is limited primarily to NOAA-9 and -14. -7 and -16 contribute a bit too, but -9 and -14 are the big winners.

    Cleaning up the script right now and I’ll send it your way. Then it’s just a matter of running all the variants and typing a bunch of gobbledygook that an editor will want to publish.

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 )

Google+ photo

You are commenting using your Google+ 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 )


Connecting to %s

%d bloggers like this: