From the work of several different people I have been able to recreate the reconstruction of the Satellite temperature for the Antarctic. From my previous work we learned that it was possible to recreate the results of the AWS trend by aligning only 3 pc’s calculated by RomanM and placed in a matrix where missing values are infilled by RegEM. I then used the same technique to recalculate trends by the re-gridded data as presented by Jeff C.
First the satellite Principal Component reconstruction.
Don’t let the terminology scare you, scientists sometimes count on that trepidation. PC’s are just curves multiplied times a coeffecient and added together to create a new trend. If you do it 55o9 times you can create a different trend for 5509 points describing the temperature of the antarctic. The process used above was to take the graphs on the right side and delete the pre-1982 data. The graphs were placed in a matrix with all of the manned surface station data as ‘originally’ presented by Dr. Steig. RegEM was run on line and the pre 1982 data was recreated looking almost exactly like the original 3 pc’s. This demonstrates a reasonable verification of the method, we can actually recreate the part of the satellite reconstruction we have data for.
I’m doing my best to ignore the unnatural unreasonable look of PC3, it’s not that easy for me.
Above is a plot of the various surface stations in the antarctic used to reconstruct the historic data for the AWS stations.
Jeff C, noting the heavy weighting of the stations in the tip of the peninsula which didn’t seem to me be reasonably addressed in the paper, came up with an idea to regrid the data on a triangular matrix which looks like this.
Temperature stations grouped into cells were turned into an individual high accuracy trend rather than reused a dozen times in the same reconstruction. The idea is to reduce the likelihood that a certain pattern would be overweighted in the trend. While the re-gridding is imperfect, IMO it is a serious improvement over the original paper.
I requested Jeff’s regridded data, which he politely provided and used it in RegEM to redo the satellite reconstruction of temp trends. The three pc’s are below.
The graphs on the left are the originals from Steig09 the ones on the right are the reconstructed pc’s using Jeff’s data. I don’t pretend to understand the reason for every detail of the reconstruction but it seems to me that the difference between the unnatural PC3 and the regridded much more reasonable PC3 is created by excessive correlation of a large percentage of peninsula series. Other than that the 6 pc’s are reasonably similar.
The original satellite data looks like this.
Now my reconstructed trend demonstrating again that I am able to recalculate Dr Steig’s work.
And the improved trend from Jeff C re-gridded results.
The slopes in the above graphs are in degrees C/year. The slope 0.007 is half the original trend and it happens to match the AWS surface data reconstruction. It seems like the regridded data which is more in filled has a more stable result than the original paper. It also is verification that station positioning (density of peninsula measurements) affects the satellite total reconstruction as well.
This graph shows the original trend distribution. Code by JeffC.
The next graph shows my own RegEM reconstructed version of the original.
Here is a graph of the regridded trends.
The final result shows a slightly different distribution of trend but a much lower overall trend indicating the sensitivity of RegEM to the weighting and density of temperature stations. These effects were apparently ignored and left un-investigated by the authors, however they are real an create significant differences in the reconstruction. After demonstrating that I can reproduce the results from the mid level pc analyzed satellite data (still not the original satellite data) and viewing the same halving of trend that Jeff C got on the AWS data by simple re-gridding, substantially more focus needs to be directed toward quality control of station weighting in the final trend.
the Air Vent 
I’m not an expert, but I can appreciate the research and the work that has gone into these analyses. Impressive.
Jeff, it looks like the trend difference may arise from the change in the PC3 trend. What are the trends in the various PCs? I’ve also noticed that most of the trend arises from very early low values – if you do trends from 1965 on, some variations have no trend.
Jeff, what is the significance of the slope and error bars/CI’s? Compared to Steig’s, as well?
#1, Thanks, this one was a bit painful because there was a ton of verification involved.
#2 I think you’re right. What you might like is that there seems to be a de-centering issue in to the reconstruction. That should be what I look at next.
#3 The paper gives some basic values like 0.12 +/- .07. I think the +/-0.07 number has no real meaning in this case as it simply captures the natural variation in the signal rather than the possible variation in the method.
Jeff, congrats once again.
0.007C/year gives us 0.7C/century- roughly in line with that over the rest of the planet.
In a way it verifies what the Global warmers are (now) saying that Antarctic temperature trends are not behaving differently from the rest of the World.
On the other hand 0.7C/century is nothing to panic about and, in any case, may be an artifact produced by the start point of the reconstruction.
All the best and power to your PC!
Great work Jeff (as well as Jeff C, RomanM & the other “co-authors” involved in de/re-constructing this study).
I will just reiterate my suggestion that no single station’s temperature measurements should be used as input to RegEM, because of the possibility of spurious correlations with bad data. A much better idea is to average several stations together that are spatially close to each other – although what criteria you would use is difficult (perhaps just larger grids?).
@Kondealer
I think Jeff would say this this regridding is JUST A FIRST STEP in correcting/improving Steig’s analysis, and further improvements may increase or decrease that 0.007C/year.
In #4, you bring up a good point. Just what is this +-0.07? When I read the paper, I thought I knew. After your analysis (and the others), I am not so sure. In fact for myself, if I read 0.12 +-0.07 as 0.05 to 0.19 is the answer, then I better go get some more data or do a better analysis. Now I question just where did that number come from. From my point of veiw, if the results of the re-analysis continue as so far, the relation of the answer and the CI will decide the value of this paper (IMO).
My two cents worth and basically a reply to #6 and quoted below.
“I will just reiterate my suggestion that no single station’s temperature measurements should be used as input to RegEM, because of the possibility of spurious correlations with bad data.”
This is exactly what the regridding accomplishes and #6 hit it right on the head. Whether or not the method used was the best, it shows that the paper produced by Steig has some flaws, potentially serious, that were not discussed/analyzed. In a previous post, I mentioned that non-gridded PC3 in graph 4 indicates an issue with data density, either there is not enough data prior to 1982, or pre-1982 PC3 shows fit error for post 1982 and is a residual component. Now, by regridding, the data density pre and post 1982 is quasi equal. Because of this, all three PC’s are now required for reconstruction. Regridding also adds additional variables to the RegEm calcs which minimizes the probability of locking onto particular data sets, thus minimizing the potential for skewed results.
It is still very important to get the satellite data to determine the quality of fit with ground data as this will also skew the end results. Increasing grid density will also help, but has its own unique problems and yields less of a return as density increases. We have seen in the past how data can be added or removed to provide and prove pre-concluded conclusions. So, for Steig, I would have to conclude (i) either he knew this and thus the result would have been against the grain of the AGW ideology and did not publish or (ii) the people that are actually reviewing the paper through this blog and many others are more analytical and caring about the truth, no matter what it is! IMO
Keep up the good, work Jeff. Now if only you could get the satellite data, a bunch of you could co-author and actually produce a respectable, un-biased and accurate report.
Sorry one last thought. What would happen to the trend if the peninsula data was completely removed?
One only has to look at PC3 to tell that something is seriously out of whack. One only has to look at the re-calculated PC3 to conclude that improper weighting of the peninsula data is the cause of the problems with PC3. Jeff, I am no climatologist but I think you (and Roman, Steve, Jeff C., and Ryan O) have connected enough dots to show now that Dr. Steig has a broken paper. No one following this in good concience can suggest that your diligence (let’s call it “team #2”) has been anything but a brilliant, above board piece of forensic work. Please, Please make plans with the other team #2 members to publish a paper on this when sufficient analysis has been completed. I think that such a publication would be that much more important and impressive considering lack of access to Dr. Steig’s code.
#9 I agree that another analysis with new PC’s excluding the peninsula altogether would likely give the most accurate reconstruction of the continent. It may not even need to be grid-weighted. Would there even be enough data to yield meaningful results?
Jeff – Nice job, the similarity to the trend in the AWS gridded reconstruction is noteworthy. You can’t see it in your grid figure up there, but there are two additional cells around 165 longitude, 53 latitude (cells U and V). Steve Mc has an interesting post up, one of the takeaways was that these two islands appear to be fairly well correlated with the peninsula. As I commented over there, both of these islands are more than 300 miles past the iceberg limit and presumably would be poor indicator of interior continental climate.
I had commented on the AWS grid thread that removing the peninsula didn’t make a significant difference in the trend, probably since the gridding had already reduced its contribution. I’ll see what happens without the islands.
@10 (Layman Lurker)
Removing the peninsula entirely would be BAD, because you are then on-purposely be removing data that RegEM can potentially find valid correlations with. That means you are TRYING to stop RegEM from predicting warming, which is just as bad as what the Team do (except more obvious than their methods).
Just remember that RegEM will ignore the peninsula if there is no correlation with the rest of the Antarctic. As long as the data fed into RegEM is good (and I do still question that), and you are careful about the RegEM parameters, then I believe it should produce relatively reasonable results.
Although I & others are still uneasy about the applicability of RegEM to generate vast swathes of data from before the satellite data begins (which seems more akin to extrapolation than interpolation – and extrapolating very far is usually a very bad idea).
This is off topic, but it looks like someone has hacked into Climate Audit.
No, I think Lucia pointed out it was a bad wordpress update.
#14 Hope that’s the case – the little skier figure superimposed on the comments section made me think the damage was something intentional.
Chris, I’m not suggesting that a run excluding the peninsula is the only case that should be presented to critique the paper. However, it is only 5% of the land mass, and if climate on the peninsula and southern islands like Macquarrie and Campbell are shown to be distinct from the greater continent, then the case excluding the peninsula should an presented as one angle of a comprehensive argument.
re: #16
Sorry, that should read “should be presentend as one angle of…”
Jeff – Dr. Steig also did another reconstruction that used plain-old PCA instead of RegEM. According to the SI, this recon infilled missing data using the monthly mean for the station. It also only used 15 stations instead of 42. Of the 15, only 2 (Faraday and Signy) are on the peninsula or northern islands. Compare that to 20 out of 42 in the full regEM reconstruction. Here are the 3 PCs, PC3 looks back to normal.
Unfortunately, both the weighting/number of stations and the infilling method changed, so we can’t be certain which caused it.
re: #18 Jeff C.
Do you have any idea why only 15 stations would have been used? Do you know how the reconstruction was used in the paper?
#19 No, I’m note sure why he used only 15 on the PCA reconstruction. According to the SI, they did the following seperate recons (#1 was the primary, the others were done to verify “robustness”):
1) Satellite and 42 surface stations using RegEM – this is the primary reconstruction (ant_recon.txt on Steig website)
2) Satellite and 15 surface using RegEM – not on Steigs website
3) Satellite and 15 surface using monthly means for infilling with traditional PCA (ant_recon_pca.txt)
4) Satellite and 42 surface using RegEM but with satellite data linearly detrended – I have not studied this and am not sure exactly what he means (ant_recon_detrend.txt)
5) AWS and 42 surface using RegEM (aws_recon.txt)
I’ve been thinking that since we have the reconstructions for both #1 and #3, and we also have the 42 station data (and know how to infill using both methods), there may be a way to work backward from both to arrive at the common satellite data. If anyone has any thoughts or ideas, let me know.
#20
If the reconstruction #3 was used to verify that reconstruction #1 was robust without most of the peninsula data, then is it likely that the infamous “Harry” station would come into play?
#21
I don’t think Harry was used, only manned station data for the 15. However, in discussing the satellite cloud masking algorithm, Dr. Steig states that any satellite measurements that did not agree with +/- 10 deg C of the climactic data were removed from the data set. He does not specify the source of the climactic data, but I wonder if Harry was involved there?
Yes I see now from your post and should have recalled that Harry was AWS. Your cloud masking algorithm comment is an interesting thought. How was the algorithm incorporated into the reconstruction?