Image above courtesy Dr. Judith Curry
The Liu and Curry (2010) paper has been the subject of a number of posts at Watts Up With That over the past few days. This post should complement Willis Eschenbach’s post Dr. Curry Warms the Southern Ocean, by providing a more detailed glimpse at the availability of source data used by Hadley Centre and NCDC in their SST datasets and by illustrating SST anomalies for the periods used by Liu and Curry. I’ve also extended the data beyond the cutoff year used by Liu and Curry to show the significant drop in SST anomalies since 1999.
Preliminary Note: I understand that Liu and Curry illustrated the principal component from an analysis of the SST data south of 40S, but there are two primary objectives of this post as noted above: to show how sparse the source data is and to show that SST anomalies for the studied area have declined significantly since 1999.
On the Georgia Tech on: “the paradox of the Antarctic sea ice” thread at WUWT, author Judith Curry kindly linked a copy of the paper in manuscript form:
http://www.eas.gatech.edu/files/jiping_pnas.pdf
Liu and Curry use two Sea Surface Temperature datasets, ERSST and HADISST. They clarify which of the NCDC ERSST datasets they used with their citation of Smith TM, Reynolds RW (2004) Improved Extended Reconstruction of SST (1854-1997). J. Clim. 17:2466-247. That’s the ERSST.v2 version. First question some readers might have: If ERSST.v2 was replaced by ERSST.v3b, why use the old version? Don’t know, so I’ll include both versions in the following graphs.
Liu and Curry examine the period of 1950 to 1999. Sea surface temperature data south of 40S is very sparse prior to the satellite era. The HADISST data began to include satellite-based SST readings in 1982. Considering the NCDC deleted satellite data from their ERSST.v3 data (making it ERSST.v3b) that dataset and their ERSST.v2 continue to rely on very sparse buoy- and ship-based observations. ICOADS is the ship- and buoy-based SST dataset that serves as the source for Hadley Centre and NCDC. Figure 1 shows typical monthly ICOADS SST observations for the Southern Hemisphere, south of 40S. The South Pole Stereographic maps are for Januarys in 1950, 1960, 1970, 1980, 1990 and 2000. Since I wanted to illustrate locations and not values, I set the contour levels so that they were out of the range of the data. I used Januarys because it is a Southern Hemisphere summer month and might get more ship traffic along shipping lanes.
http://i37.tinypic.com/x1wtvm.jpg
Figure 1
As you can see, there is very little data as a starting point for Hadley Centre and NCDC, but they do manage to infill the SST data using statistical tools. Refer to Figure 2. It shows that the three SST datasets provided complete coverage in 1950 and 1999, which are the start and end years of the period examined by Liu and Curry. For more information on the ERSST and HADISST datasets refer to my post An Overview Of Sea Surface Temperature Datasets Used In Global Temperature Products.
http://i34.tinypic.com/j5jhp5.jpg
Figure 2
A question some might ask, why did Liu and Curry end the data in 1999? Dunno.
As noted above, Liu and Curry illustrate data for the latitudes south of 40S. There are differences of opinion about what makes up the northern boundary of the Southern Ocean. Geography.com writes about the Southern Ocean, “A decision by the International Hydrographic Organization in the spring of 2000 delimited a fifth world ocean – the Southern Ocean – from the southern portions of the Atlantic Ocean, Indian Ocean, and Pacific Ocean. The Southern Ocean extends from the coast of Antarctica north to 60 degrees south latitude, which coincides with the Antarctic Treaty Limit. The Southern Ocean is now the fourth largest of the world’s five oceans (after the Pacific Ocean, Atlantic Ocean, and Indian Ocean, but larger than the Arctic Ocean).”
But isolating the Southern Ocean for climate studies really isn’t that simple. The Antarctic Circumpolar Current (ACC) is said to isolate the Southern Ocean from the Atlantic, Indian and Pacific Oceans. Unfortunately, the northern boundary of the ACC varies as it circumnavigates the ocean surrounding Antarctica. Refer to the University of Miami Antarctic CP current webpage.
In this post, I’ll illustrate the SST anomalies of the area south of 40S that was used by Liu and Curry. They capture additional portions of the ocean within the Antarctic Circumpolar Current. (They also capture small areas north of the ACC.) And I’ll identify that data as the Mid-to-High Latitudes of the Southern Hemisphere (90S-40S).
I’ll also illustrate the SST anomalies of the Southern Ocean, as defined above (south of 60S), because they capture the Sea Surface Temperature anomalies of the Southern Ocean most influential on and influenced by Sea Ice. Let’s look at that data first.
THE SOUTHERN OCEAN (90S-60S) SST ANOMALIES
Figure 3 compares the three versions of Southern Ocean (90S-60S) SST anomalies, from January 1950 to December 1999, the same years used by Liu and Curry. Included are ERSST.v2, which Is used in Liu and Curry, ERSST.v3b which is the current version of that dataset, and the HADISST data, also used in Liu and Curry. All three datasets are globally complete. And as shown in Figure 1, the Hadley Centre and NCDC have to do a significant amount of infilling to create spatially complete data for those latitudes. The data has been smoothed with a 13-month running-average filter to reduce the noise. Also shown are the linear trends. Again, this is not the full area of the Southern Hemisphere SST data used by Liu and Curry. I’ve provided it because it presents data that is more impacted by (and has more of an impact on) Sea Ice. The linear trend of the ERSST.v2 is almost twice that of the HADISST data. Note also the change in the variability of the HADISST data after the late 1970s. HADISST has used satellite data since 1982 and this helps capture the variability of the Southern Ocean SST anomalies.
http://i36.tinypic.com/287ejkg.jpg
Figure 3
Figure 4 shows the Southern Ocean SST anomalies for the ERSST.v2, ERSST.v3b, and HADISST from January 1950 to December 2009, with the data smoothed with a 13-month filter. The HADISST data peaked in the early 1990s and has been dropping since. This was not easily observed with the shortened dataset. The two ERSST datasets peaked in the early 1980s. For all three datasets, the recent declines in the SST anomalies have caused their linear trends to drop sharply from the values presented in Figure 3. In fact, the HADISST is now basically flat.
http://i36.tinypic.com/snea10.jpg
Figure 4
MID-TO-HIGH LATITUDES OF SOUTHERN HEMISPHERE
Figure 5 is a comparison graph of the SST anomalies for the latitudes (90S-40S) and years (1950-1999) used by Liu and Curry. Note how there are two distinctive periods when there are sharp rises in SST anomalies: from 1966 to 1970 and from 1974 to 1980. Then from 1980 to 1999 the SST anomalies for the mid-to-high latitudes of the Southern Hemisphere flattened considerably. The HADISST data flattened more than the ERSST datasets.
http://i34.tinypic.com/4kwc51.jpg
Figure 5
In Figure 6, I’ve included the data through December 2009. Note the significant drops in the SST anomalies in all three datasets. All three peaked in 1997 (curiously before the peak of the 1997/98 El Niño), and have been dropping sharply since then.
http://i37.tinypic.com/3447dk7.jpg
Figure 6
CLOSING
The title of Liu and Curry (2010) “Accelerated Warming of the Southern Ocean and Its Impacts on the Hydrological Cycle and Sea Ice” contradicts the SST anomalies of the latitudes used in the paper. The SST anomalies are not warming. They are cooling and have been for more than a decade.
SOURCE
The Maps were created using, and the data is available through, the KNMI Climate Explorer:
http://climexp.knmi.nl/selectfield_obs.cgi?someone@somewhere
Dr Curry in a recent interview:
“Some people were getting their papers rejected because they disagreed with the IPCC.”
Sounds to me like something worth pressing her on, the implications if true are astonishing.
Dr. Curry and Dr. Liu; I thank you for this venture in collaboration and especially your determination to see the venture through!
As for your paper; I heartily agree with John Whitman’s summary and Tom Vonk’s commentary and the opening posts of Bob Tisdale and Willis Eschenbach. Why?
I tried (and am still trying) to separate the paper into assumptions, observations and findings. I expected firm statements somewhere in the document defining each category clearly. Instead I had to work hard in chasing down terms, where they came from and what they meant. Because of this, the paper comes across to me as a very parochial document meant for distribution within a close knit group. A group where terms are exchanged so frequently that they become familiar and their real meaning is “understood” by insiders. No, not climate science researchers; more like a subset group within climate science. One example of this is the use of the term snowfall in the document. Snowfall as used is a subset of the increased precipitation described. Only by chasing down all references to snowfall and precipitation did I determine that snowfall is derived from a model (20C3M of CCSM3) and (A1B of CCSM3). Silly me! I was assuming that snowfall was an observation; nope it is an assumption.
There are some really confusing statements, again perhaps because the document is meant for distribution within friends. A real concern exhibited by Willis and seconded by many people concerned the data quality prior to the 1970s. Dr. Curry explained this away by stating that she had confidence (my words) in the 1950-1999 data. OK? Well, no! Once I started trying to follow ideas and thoughts through the document I kept getting struck in the face by the lack of information about the specific years used by ALL model runs. Another assumption, I suppose, but one that leaves me uncomfortable. This discomfort level is heightened by the use of PIcntrl. What the freak is a pre-industrial simulation doing in a study where data before the 1970’s is dismally poor at best? Is this because data is so sparse that a simulation is needed to buttress the data infilling assumptions and show an unnatural warming? I don’t know, or I dunno as others have pointed out.
I am really puzzled why pre industrial model runs were used for the 1950-1999 period but no one involved with this document bothered to even a mention recent 2000-2010 observations as confirming, refuting or even that the hypothesis needed further study.
My understanding of science may not be your understanding. When hearing proposal for a hypothetical rationale to a (mostly) computer generated problem, I personally would be more interested in proofs and the methods to obtain those proofs. Instead I spent two days deconstructing sentences in subject lines, modifiers and qualifiers and finally figured out that assumptions were made, models were built and simulations ran using data that should not have been used.
I am also confused by the chosen title for the paper. The term accelerated, does that originate from the GCM models? As far I can find in the document this acceleration is caused by increased GHG in the GCM models. Warm begats warm waters right? So your model uses GHG caused exponential warming as a base assumption. What I find really confusing is the second part of the title “Its Impacts on the Hydrological Cycle and Sea Ice”. Based on that second title line my assumption would normally be that this was a fact finding study detailing the effects caused by warming to date. My bad, I couldn’t have been more wrong in that thought.
Which brings us to;
1. Judith Curry says:
August 18, 2010 at 8:18 am
“…Our paper compares model simulations with available observations (we consider two different data sets) in an effort to unravel the physical mechanisms that determine Antarctic sea ice extent in response to climate variability and change.
We have identified a plausible physical mechanism that seems to make sense. Science is about trying understand how things work.
We have made no extravagant claims in either the paper or the press release. …… It talks about the increase of Antarctic sea ice, which is hardly a talking point for alarmists
Yes, climate models are imperfect and there are deficiencies in SST data sets particularly in the first two decades of the period that we examine. So we have imperfect tools to test our hypothesis. Others will examine this problem from different angles. Eventually we will have better data sets and better models to work with. That is how science works.
This paper raises an issue that climate researchers should pay more attention to. Since the climate model simulations of Antarctic sea ice generally agreed with observations, climate researchers would say “consistent with” without really understanding the mechanism. And we definitely need more and better data in the Southern Ocean…”
I doubt that truer words could ever be spoken in climate science and we should certainly respect that a co-author would know better than us the pitfalls inherent in the study. Honesty with us is terrific, honesty with the world would have been better. Your simple summary should be in the opening statement and parts of it repeated in the findings.
From a personal perspective; if I was handed a paper written like this to review (in any discipline) I would’ve redlined large sections and sent it back for additional work. Papers should clearly introduce the audience to the paper, explain and document the data, tools, methods and findings then exit on a succinct statement that leaves the reader satisfied they grasped the content. This goes even for documents targeting specific audiences. Again, I emphasize that this document assumes very reader is an inner circle initiate. Please, please, please; explain terms and phrases that are not general understanding. Snowfall indeed, I’m thinking white crystalline stuff and the writer is thinking simulated millimeters.
I hope you have a wonderful and productive trip Dr. Curry! I’m looking forward to your return and future participation.
Jason
That was a telling quote. Here is the interview with Judith Curry.
She’s a brave person and I respect her courage in speaking out. I’m not impressed with the practice though of writing papers using data that doesn’t exist or is so sparse its almost invisible
http://blogs.chron.com/sciguy/archives/2010/08/judith_curry_on_antarctic_ice_climategate_and_skep.html
For the layman, like me, I suppose the best description of this paper is that it is not a “What is” study but a “What if.” The media, not surprisingly, considering its title, chose the What is.
When I look at the plots of the data, a linear fit isn’t what jumps into my mind.
If one fits a nonlinear function with a linear function, then the cure is a lagging indicator.
That is, the trend is looking backwards – the curve has no predictive power.
Where is the original data?
Smoothed experimental data without an indication of the uncertainty in the data is junk science.
Judith, I am a fan of yours. When you enter a discussion, it becomes much more professional and the data flows readily. It reminds me of my years spent in industry wrestling with problems and situations that had never been addressed before. Without bias and politics, technical problem solving is much more fun and exciting. Liu also entered this discussion so he is to be commended as well.
The fact that I have sat you on a personal pedestal wearing a golden crown, does not impact my professional opinion of your paper, however. You got your technical butt kicked hard and often by some real pros, led by Willis Eschenbach and Bob Tisdale. Your paper raised many, many good rebuttals and a bit of heat, but your excellent standing in WUWT and a bit of nudging from Anthony, kept the rebuttals on a high plane. I am sure that you know without me saying that you are highly regarded and respected in WUWT.
May I offer a suggestion for you and Liu? Pull your paper, carefully read the three discussions of your paper, find a quiet time and go sit on a stump in the woods and reflect on the discussions, then do the mini-studies suggested and finally draft a follow-up paper detailing the overall findings basis the new studies and insights.
Next, privately ask Willis, Bob, Anthony and a few others that offered meaningful rebuttals to your first paper, to review your draft. Revise the paper in light of these reviews and start the publication process over.
And watch the title!
JFD
JFD
I would like to endorse all the comments and sentiments in your post.
tonyb
I’m taking credit for this. I haven’t seen it before, but I think it can very easily be summed up as GMIGO: “grant money in, garbage out”!
TomVonk
You raise interesting points. I’ve no doubt that if someone took a very serious look into the EOF analysis (on this dataset) they would find several ways to rip it to shreds.
A problem that arises with PCA, EOF, & factor analyses more generally:
People not only don’t do thorough diagnostics – they don’t even know how – worse still: they don’t even know they should. The preceding comment isn’t aimed at anyone in particular. It is an observation arising from hanging around both stats departments at universities & stats consultants.
I also strongly object to defining the Southern Ocean as 40-90S. Even just plain 60-90S is a simplification. For preliminary investigations I recommend using the Antarctic Convergence and tectonic plate boundaries as guides for spatial analysis of the Southern Ocean – (the latter will include the Southeast Pacific & a Humboldt Current component).
I agree very strongly with Dr. Curry that anything that stimulates research on the maritime deep south (not Antarctica alone) is both welcome & due. (I would include the Southeast Pacific beyond just 60-90S.) The excessive focus on the northern hemisphere (and even on the northern portion of the southern hemisphere) is unacceptably smothering.
Continental patterns differ dramatically from maritime patterns. This is not something that simply splits symmetrically over the equator. The ‘dividing line’ does not even follow a line of latitude. (See definitions suggested in the previous paragraph for preliminary investigations of the Southern Ocean — investigators like Bob Tisdale help refine our focus futher…)
Correction:
not “tectonic” — rather, this is what I had in mind:
Earthquake Map:
http://earthquakes.usgs.gov/research/data/plate15.pdf
Using the Southern Ocean / Southeast Pacific boundary suggested by that map (for bounding SST geographically), one finds interesting coherence with stratospheric aerosol optical thickness – (raises more questions than it answers – exactly what makes it interesting).
Bob Tisdale
Any ideas of how to present this for readers without science backgrounds?
Yes .
The most basic notion and this one is necessary for all PCA , EOF etc methods is the notion of coordinates .
So you always must begin by a presentation that there is an infinity of possible coordinates among which you choose arbitrarily one .
Then I generally follow by a rugby ball .
If you choose first a coordinate system and then put a rugby ball in those coordinates , then its description is complicated . There are no symmetries and you might actually be unable to recognize that you have a rugby ball .
So the idea is to pick among the infinity of coordinate frames a unique special one which is chosen by observing the symmetries of the rugby ball .
So clearly if you choose the symmetry axis of the ball as 1 direction and a plane orthogonal to the axis as 2 more directions , you obtain a new frame of reference where the rugby ball description is obvious and simple .
The projection of the ball on the orthogonal plane are just circles and its projection on a plane containing the symmetry axis are just ellipses . The rugby ball can be recognised in this frame by simply looking at it (hence the notion of “recognizing the needle” in my post above) .
Now when you collect some data and visualise them in some arbitrary coordinate frame with N dimensions , those data points will just be some N dimensional rugby ball . What the PCA , EOF etc method does , is to find a new coordinate system in which the rugby ball becomes obvious . F.ex the coordinate axis which goes along the most stretched direction of the rugby ball is the most significant (technically “has the highest eigenvalue”) and will show the best correlations .
So its all just about a method of changing an arbitrary coordinate frame in a special one.
Of course if your data don’t really form a rugby ball but a badly deformed potatoid , then changing the coordinate frame doesn’t help much . It still stays a badly deformed potatoid and you recognize nothing .
.
Atheok
Yes , that is what I tried to say in my post . The biggest issue with this draft is that it is impossible to tell what is data and what is just some computer programm .
This is most striking in the P-E case (Precipitation – evaporation) . It’s already bad enough that the link given in the Ref doesn’t work . But it takes unholy time to realize that these numbers are NOT data but again some computer simulation .
Actually when you look at the figures there are only THREE that deal with real data – 1a , 2a and 3a . Everything else , that is 27 figures , are just computer simulations !
That means that the draft is dedicated to 90% to the behaviour of computer models without saying that it is so .
The summit is reached in the 1j , 1l and 1m figures .
In those figures all structure is destroyed , the Antarctic is not only represented by EOF1 , it IS EOF1 .
That means that the phase space became monodimensional !!
For the computer the whole of Antarctic can be explained by only ONE variable and as the domain is a homogeneous red blob , the relationship is linear .
What is this variable ? Well CO2 concentration because it is the only thing that varies between Fig1a (the reality) and Fig1l (the computer) .
The CO2 has overwhelmed and crushed all non linear dynamics in such a way that for the computer it became the only thing that matters – forget ENSO , PDO and all complex spatial structures .
This is clearly not what happens in the real world .
Strangely the authors instead of saying that the computer behaves in an unrealistic way and that the results should be rejected , seem to treat the fact that the phase space became monodimensional as having as much credibility as real (measured) data that obviously contradict it .
This is a much too common feature of “climate” papers . People simply stop making a distinction between data and computer runs . You have no data about snow fall ?
It’s not a problem , you run a computer model 100 times , average and voilà ! You have all the “data” you want . This is not good science in my opinion .
.
Paul Vaugn
I’ve no doubt that if someone took a very serious look into the EOF analysis (on this dataset) they would find several ways to rip it to shreds.
I agree . I strongly doubt that the draft will stay in the present form .
I am pretty sure that there is a sampling problem because the first eigenvalue is much too low . That means that the second is not far from teh first and/or that the eigenvalues are closely spaced .
I am also pretty sure that the size of the domain is too big . That means that the classical test of cutting the domain in 2 and running EOF on each of the 2 parts (I mean here the real data and not computer simulations) will show a significant sensibility to the domain definition .
And if it does , it is enough to destroy the whole idea .
Well just as an aside, I just (this minute) bounced EOFs off a high powered PhD Mathematician (Indian chap) who sits right adjacent to me. He spends his whole day wrapped up in analysis and filtering of signals of all manner; has Morlet Wavelets for breakfast.
He nearly had a heart attack on the mention of “empirical” with relation to “orthogonal functions” and pleaded complete ignorance (like me).
So I guess I’m in good company; in that he is very good at what he does; but evidently both of us have been leading charmed lives to have never run across EOFs.
He wiki’d it up and took a look; and confirmed that it was not within his ken; but that he could probably get up to speed PDQ (I’d agree on that (for him; not for me)).
So you blokes who use it routinely are out there in hyperspace where some of us have never been.
But when my colleague gets aboard; he will probably be able to explain it to me.
In the meantime; am I correct in that EOFs are NOT a method of synthesizing some arbitrary function in a representation that permits recovery of the original function, without loss of information; but may bring insight into the arbitrary function that was note previously perceived ??
It was the apparent replacement of a host of information, with a dearth of information in another form; in apparent contradiction of the Nyquist Sampling Theorem, that had my alarm bell ringing.
Thanks, Anthony, and thanks, Judith Curry.
George E.Smith
In the meantime; am I correct in that EOFs are NOT a method of synthesizing some arbitrary function in a representation that permits recovery of the original function, without loss of information; but may bring insight into the arbitrary function that was note previously perceived ??
The short answer is yes .
As I have repeatedly written on this thread , EOF is PCA for all practical purposes .
I don’t know the origin of the “EOF” name and it is rather irrelevant . It established itself in geophysics when scalar spatiotemporal fields were analysed (temperatures , pressures , densities) .
The purpose is to find spatial patterns (standing waves) that capture the variability in the data .
So to use your vocabulary we have some unknown function f(x,t) that has been measured . Then a matrix M(xi,ti) is formed where the columns are locations and the lines times . For example 1 column is the time series of SST in Ushuaia . Etc .
You remove then generally the temporal mean from each column and form the covariance matrix .
The EOF method will give you spatial variability modes (e.g EOFs) which explain most of the variance in the data .
Normally a graphical represenattion of an EOF is a contour map that tells you how much of the local (measured) variance is explained by this considered EOF .
At one place you may see 0.9 what means that this particular EOF explains almost all variability at this place .
At another place you see 0.1 what means that this particular EOF explains almost nothing of the variance at this place .
Let us note that this is not what the Figures in the paper show even if they should .
Of course you can reconstruct all original data of the covariance matrix , so f(x,t) , from all EOFs and their Eigenvalues .
However as the purpose of all PCA like methods is to simplify , you will always want to truncate .
I don’t know how many eigenvectors (EOFs) there are in the matrix considered here .
Let’s suppose 50 .
As you see the authors kept only ONE EOF which explains one fourth of the variance .
Clearly you can’t reconstruct the data with this one EOF because there are three fourth of the variability which are NOT explained by this EOF .
As I have already written in previous post , the purpose of PCA/EOF is to identify spatial modes (as they are orthogonal , they are spatially uncorrelated with each other) which explain a big part of the variability .
So no , you don’t get insights in the f(x,t) itself but you get insights in its spatial modes of variability .
This is the easy part .
The hard part begins when you try to connect a contour map of an EOF (I remind that the components of the EOF vector are generally correlations e.g dimensioneless coefficients ) to a physical process .
This is what von Storch called “recognizing the needle by just looking at it” .
If you can’t do that , then any PCA/EOF method is physically useless even if it is mathematically correct .
I hope it helps .
TomVonk: Thanks.