Thursday, March 24, 2011

Trends in Antarctica

In a comment, Eric Steig suggested direct estimation of temperature trends in West Antarctica(WA) (as opposed to estination via fitted EOFs). For trend estimation, the methods of S09 and O10 have good and bad points. A plus is that they do estimate for a prescribed area. Against that is that it isn't obvious, with variable sea ice, what area to prescribe. A minus is the indirectness. Temperatures are used to fit basis functions (EOFs), and the trend is obtained from the fitted function. Whether through a small number of EOFs, or via regularisation (kgnd), this interpolated step restricts degrees of freedom.

TempLS is designed for the direct estimation of trends via least squares model fitting, so I thought I would try that direct approach. I deferred the project while I looked for a suitable area weighting scheme, and that is now available.
Background to this post can be found in earlier discussions here, here, here, here, here, and here

The discussion is prompted by two papers:
S09, a 2009 Nature paper by Steig et al and RO10 (or O10), a 2010 J Climate paper by O'Donnell et al. There has been heated blog controversy about these papers - this post is just one entry point. Here is a Real Climate post following S09's original appearance, and here is Eric Steig commenting on O10 (this led to other heated postings).
There are earlier discussions of area weighting by RO10 authors at tAV, eg here and here.


The results below broadly confirm the earlier spatial analyses, and follow the regional breakdown of S09 and O10, though often with higher trends. The continent ground station trend, 1958 to 2009 and not area weighted, was 0.175 ± 0.03 °C/decade. This is consistent with my earlier EOF-based result, and higher than S09 or O10. Area weighting brought it down to 0.114 ± 0.035 °C/decade, similar to S09 (with AVNRR). Adding equally weighted AVHRR pushed it up to 0.137 ± 0.037 °C/decade. I think these results add a degree of confirmation to those earlier analyses, and to my EOF-based analysis, and tend to confirm a positive warming trend, especially outside East Antarctica.

Something that needed attention is the regional tagging of the data I am using, which comes from the data Ryan put on line, and earlier from S09. There are four regions, Peninsula, WA, Ross Sea, EA. But the satellite tags are bounded inconsistently with the ground stations, so I retagged them as shown in this map. Ground stations are shown with large dots, and the subset of AVHRR stations (described here) are small dots, both colored by region.



That done, I have computed trends for these regions and the continent using the OLS methods of TempLS V1. As with some earlier posts, I have mixed AVHRR and ground stations, with various relative weightings (including ground only).



Continent ground station trends

These are difficulties with the sparse samples in some tof the regions, most notably with West Antarctica. So the whole continent is simplest. I'll start with unweighted (by area):

Stations reporting in each year

A map of ground stations colored by region

The trend is quite high at 0.175°C/decade. Now I'll add weighting by area, as described here. The map shows for a specific month, Dec 1987,  the triangular mesh created, the subdivision used to weight each station, and the circles show, by area, the size of the weights. I have not smoothed, nor tried to cut out the Weddell Sea.




The trend is lower, at 0.114°C/decade. This is in line with earlier observations that area weighting reduces the trend, probably by boosting the effect of the less rapidly warming East Antarctica stations, which occupy the largest region.

Adding AVHRR readings


As with previous posts, I've introduced the satellite AVHRR data, which comes as grid data, as artificial stations. I've selected only 1 in 4, so there are 1392 of them. A factor is needed to determine the relative weighting for combining them with ground stations. I now use a range from 0 (all satellite) to 1 (all ground), with 0.5 indicating that total weights for satellite and ground are equal. That is for all years, so that before 1980, there is ground data only, and the factor doesn't matter. Because of the way it is totalled, this means sat data gets a higher weighting (at 0.5) after 1980. This is a small effect.

So here are results from satellite data only. Like O10, I'm using Eric Steig's archived AVHRR data, which goes up to 2006 only.

As expected, the trend is quite high. One of the reasons for interest in S09 was that it looked into the discrepancy between the ground trend (lowish) and AVHRR (high).

So now looking at the combination of satellite and ground, remembering that before 1980 it is ground only:

So the trend is lower - closer to ground than AVHRR, reflecting the shorter period of AVHRR.

Antarctic Peninsula trends

So now I look at regional trends. I'll start with the Peninsula, which does not have problems of years without data. Here are ground stations only:

AS expected, the trend is high. Now if the AVHRR stations are included, with equal (total) weighting:

East Antarctica trends

Again firstly the ground stations:


As reflected in work in previous posts (and S09, O10) ground stations in East Antarctica have been warming very slowly. The inclusion of AVHRR results at nominally equal weighting increases this by a fairly small amount:




West Antarctica and Ross Sea

I found a specific difficulty with WA, which I think Ryan and Eric have mentioned. In about 1980, about the time AVHRR satellite information is available, there was also a change to measurement (AWS). Only one station in WA was reporting before then, Byrd, and there was a gap of four years with no readings at all. Byrd changed to AWS, and this is listed as a new station. So there is no way of bridging the gap without making some assumption about how the old Byrd relates to Byrd AWS. And even that would be slender information.

A first option for WA is to combine it with the Ross Sea stations, where there is a fuzzy boundary anyway. There are just enough stations pre-1980 to get a result:

Again, as expected, the trend is high. Adding AVHRR stations does not change this.



West Antarctica trends

The other option for WA is to look at results after 1980 only. That gives:

The trend is very high indeed, but so is the uncertainty, and the stations are few.

It seems the AVHRR data stabilizes the average when it is available, so just checking those years:


This did reduce the trend slightly, and improve the error range.







32 comments:

  1. What the hell happened with the AVHRR in 1994? Is this a physical or a calibration/instrument issue?

    ReplyDelete
  2. Eli, I don't know. I've looked over the numbers for typos etc but can't see anything. From the regional breakdown, it seems to be mostly E Antarctica, but for AVHRR thats a lot of "stations".

    ReplyDelete
  3. Eli, I did some more checking. I looked, for all AVHRR, at 1994 monthly relative to the average of 1993 and 1995. Overall the mean difference was -1.35C, which looks about right (if I'd used 1992 and 1996 it would have been bigger). Monthly, the numbers were:
    -0.95 -3.79 -3.21 -0.91 -0.99 1.38 -1.32 -2.67 0.22 -1.73 -0.81 -1.37
    So cold Feb and March, but no obvious error. I made a histogram of the annual average diff over stations, but there was nothing exceptional there.

    ReplyDelete
  4. How do you know that giving a (relatively) large area weighting to coastal measurements is valid? Aren't you effectively measuring sea temperature by doing that?

    ReplyDelete
  5. Anon, one of the ways in which Antarctica is different is that "coastal" can be a long way from open sea. One of the things that bothers me is that the entity that we're talking about is ill-defined and seasonal. So one could argue that coastal stations deserve weighting for the adjacent sea ice.

    But the other aspect is that coastal is what we mostly have, and if they are downweighted, there would be dependence on very few stations.

    ReplyDelete
  6. Could you make an AVHRR - station (weighted) trend map so we can see the areas where they differ?

    ReplyDelete
  7. Nick, The temporal (particularly trend) information in the AVHRR data is very suspect. There appears to be drift in the data from individual satellites (one in particular had serious problems) and inhomogeneities on the switchovers between successive satellites. Ryan did quite a lot of work on this early on, some of which was posted at TAV I think. As I recall, the AVHRR data for grid cells in which ground stations are located generally show a substantial positive trend bias. Further, the AVHRR instruments don't measure air temperature, but rather the surface skin temperature, which is not the same thing (there is a strong temperature gradient just above the surface).

    Therefore, IMO any temperature reconstruction for Antarctica that uses the temporal information from the satellite data is suspect and likely to significantly overstate the true near surface air temperature trend.

    The spatial correlation information from the AVHRR data seems to be of much higher quality, being relatively unaffected by temporal drift and inhomogeneities. Therefore, whilst there are substantial uncertainties (principally in West Antarctica, but to a lesser extent elsewhere), I have more confidence (supported also by an extensive verification exercise) in the accuracy of the method we employed in OLMC10 than in the method you use here. You might like to try withholding surface station data (one at a time, or a set corresponding to verification stations withheld in OLMC10) and seeing how the verification statistics for the withheld stations compare with those reported in OLMC10 (which were far superior to those for S09).


    Nevertheless, I am interested in your LS approach and plan to have a play with the code that you have helpfully provided.

    ReplyDelete
  8. CCE,
    Although I'm showing maps, dividing into regions and weighting by area, this analysis isn't otherwise spatial. There was the unweighted analysis done in an earlier post. Here is ground only, and here is AVHRR only.

    I did the area weighted equivalent here. There's a range of ground/AVHRR mixes - I didn't do ground only there, but the mixing factor 5 is close. And for AVHRR, the factor 0.2, which I showed, would give a similar picture to AVHRR only.

    ReplyDelete
  9. Nic,
    I've been reading this tAV post, which may be the one you're referring to. I've added it to the reference list at the top.

    Yes, there is certainly a question of whether the AVHRR results add more signal than noise to the temporal information (there was plenty of noise anyway). You can either just use the spatial EOFs or incorporate the temporal info too. It's not clear to me how your method mitigates these drift issues, but careful reading of that post may help.

    The fact that AVHRR isn't a surface temp shouldn't matter much - the local offset absorbs those discrepancies along with altitude diferences etc.

    I'm planning to post a V2.1 of TempLS real soon. This will include the new mesh-based spatial weighting and the adaptions used to include the AVHRR (or similar) EOF's. It will also be better structured and easier to follow (I hope) - I've made more use of R lists and structuring.

    ReplyDelete
  10. Thanks Nick. I'm still foggy about the particulars of your algorithm, so I'm not sure if I know exactly what I'm asking since the mesh changes month by month. But I'm interested in a difference map (or whatever it is called) for the trends during the satellite era only. i.e. satellite trend minus station trend using your meshing algorithm (no satellite EOFs). The station data would be pretty coarse, but still good enough (I think) to show problems with satellites. If the satellites have step changes and drift problems, but are otherwise fine spatially, the difference trend should be pretty uniform over the continent, shouldn't it?

    ReplyDelete
  11. I advocated the use of cross-validation by withholding stations in an earlier post, and was pointed to the analysis at TAV. In response to another comment there was also a link to a post on spatial correlations in the AVHRR data, which showed strong correlations across long distances in Antarctica.

    Unfortunately it seems to me that the second observation makes the first invalid. If stations are strongly correlated, then withholding one and finding that it's temperature can be reconstructed from the others doesn't tell us much (the extreme case being placing two weather stations at the same site - removing one and observing the the temperature can be reproduced from the other is a meaningless test).

    Not sure how to deal with that. My gut says some sort of analysis for correlation, orthogonalisation, and then prediction. But that just sounds like the EOF calculation, and it's not clear to me what you omit. Need to think more.

    Nick: do you think the area weighted approach might be applicable to the analysis of tree-ring proxies, as an alternative to EOFs, as well?

    Kevin C

    ReplyDelete
  12. Kevin,
    I don't think area weighting is an alternative to EOF's - it tells you how important places are, but not what you should do with them. But I think it could help in putting together a global or hemispheric mean.

    On cross-validation, I think the situation you envisage is interesting. Strong correlation means that you're getting a good signal for the whole area but can't say much about regional differences. But often, why you want to figure out the regional variation is to remove it and tease out the overall signal.

    ReplyDelete
  13. Nic,
    I wasn't actually referring to my TAV post that you link to, which involves use of the MSU TLT global satellite data, although I do think that applying the RLS method to global temperatures has merits (and there is much better satellite data than TLT to use over oceans).

    Rather, I was thinking of Ryan's much earlier post at http://noconsensus.wordpress.com/2009/05/20/antarctic-coup-de-grace/, which was specifically about the Antarctic satellite AVHRR surface temperature data.

    This has a graph comparing the trends of ground stations and of their corresponding AVHRR grid cells, showing that the mean AVHRR trend is far higher, and also evidences the drift and inhomogeneity problems on satellite switchover that I mentioned.

    The AVHRR spatial correlation information is little affected by the drift and inhomogeneities, since these can be expected to affect the recorded temperatures for all grid cells similarly. As the correlations correspond to sums of monthly average temperature cross-products, they will all be slightly overstated as a result of inhomogeneities that cause a common shift in the temperature error over all grid cells. However, the jumps are small relative to the monthly standard deviations of temperature so their effect on correlations (unlike their effect on trends) is small.

    ReplyDelete
  14. Sorry, Nick, I left the 'k' out of your name!

    ReplyDelete
  15. Nick, thanks for directing me over to these newer threads on use of your algorithm with Antarctica temperature trends.

    I am attempting to keep what you doing into the perspective of the S(09) and O(10) results. You seem to be content with the higher trends that using the AVHRR data in temporal mode gives. NicL on this thread and O(10) make some rather compelling arguments against using the satellite data in the temporal mode given the potential for drift and discontinuities from satellite to satellite.

    For the sake of comparisons would it not be wise and informative, if possible, to separate the effects of the use of your algorithm from the temporal considerations for the AVHRR data. In other words, in an apples to apples comparison does your algorithm give different results than O(10) and/or S(09)reported?

    Your results show much the same pattern of trends that I noted in the early S(09) analyses and that being that most of the overall Antarctica warming shows up in the first decade or so of the series. S(09) had little or nothing to say about this split trending. Would you care to comment on that matter? Could there be a cyclical pattern in the Antarctica that would make the starting (and ending) point critical to measuring trends?

    I believe it was Steig's data that showed parts of the Antarctica were warmer in the period 1935-1945, as I recall.

    ReplyDelete
  16. Kenneth,
    As I said on the other thread, I don't think the quality of the AVHRR data is one that the analysis can resolve. That's why I've left the mixing ratio (AVHRR vs ground) as a free parameter. It's for the user to decide which to put their faith in.

    So I can't do a strict apples-apples comparison, because I have a range of results. The trend is higher for more AVHRR, but it's still fairly high for ground only.

    There will be an implied value for this parameter in S09 and O10, and I'll try to work out what it is. Then I could do that comparison.

    I don't want to make too much of the split trending. The pre-1980 data is sparse - Steig has shown how to get much better use out of it, and I think it's fairly safe to say that there is reasonable confidence in the uptrend, but identifying features within the trend is riskier. And of course, the data was even sparser in 1935-45.

    ReplyDelete
  17. Nick, before you get too complacent about the cyclical nature of Antarctica temperatures you might want to look at the paper listed below and the table in the link below. It would appear that the early 1950s was a trough in the temperature record.

    Antarctic temperatures over the past two centuries from ice cores authored by:
    David P. Schneider,Eric J. Steig,Tas D. van Ommen, Daniel A. Dixon, Paul A. Mayewski, Julie M. Jones, Cecilia M. Bitz:

    "We present a reconstruction of Antarctic mean surface temperatures over the past two centuries based on water stable isotope records from high-resolution, precisely dated ice cores. Both instrumental and reconstructed temperatures indicate large interannual to decadal scale variability, with the dominant pattern being anti-phase anomalies between the main Antarctic continent and the Antarctic Peninsula region. Comparative analysis of the instrumental Southern Hemisphere (SH) mean temperature record and the reconstruction suggests that at longer timescales, temperatures over the Antarctic continent vary in phase with the SH mean. Our reconstruction suggests that Antarctic temperatures have increased by about 0.2°C since the late nineteenth century. The variability and the long-term trends are strongly modulated by the SH Annular Mode in the atmospheric circulation."

    http://ff.org/centers/csspp/library/co2weekly/20061013/20061013_02.html

    Also, Nick I see how S(09) and O(10) were able to use the satellite data (spatial relationships) in the pre-satellite time period, but I continue not to see how you connect the pre and post satellite records with your algorithm.

    ReplyDelete
  18. "I continue not to see how you connect the pre and post satellite records with your algorithm."
    In basically the same way as O10 and S09. I use the EOF's (as calculated by O10 and S09) from the AVHFF 1980-2006 to spatially interpolate the ground data from 1957, month by month. They are the spatial basis functions. The presumption, as with S09 and O10, is that the same spatial patterns are useful pre 1980.

    They are also used for fitting the AVHRR data itself, when mixed with ground. Of course they were constructed to be LS optimal for that.

    Thanks for the info about annular modes etc. What I'm doing is a fancy version of linear regression. It doesn't look into underlying causes. A more complex model could allow for annular modes and other postulated patterns.

    ReplyDelete
  19. Nick, thanks for the info on the use in your algorithm of the spatial relationships of the AVHRR pre 1980. That it is the same process as used in O(10) and S(09) helps when looking for explantions of differences in results between those methods and yours.

    Do you have a link to the R code used for the Antarctica calculations? I always put off wading through code, but it is the best way to understand what a method is doing.

    ReplyDelete
  20. Kenneth,
    Yes, the R code is V2.1 of TempLS, which I have just posted. There is a zip file on the doc repository, TempLSv2.1.zip, and as well as the main code it includes (in userfiles.zip) two files, Antarctica.r and Ant_Area_wts.r. These contain the setup info for the first two posts.

    To run the first, you just type, in R, jobname="Antarctic" and then source("TempLSv2.1.r"). You'll need to ensure that eofs.txt (with the AVHRR eofs info) and ryan.r (with Ryan's special Antarctic graphics routines) are available. Results will be in a subdirectory ./Antarctic.

    ReplyDelete
  21. typo, the first file is called Antarctic.r. To run the second, just jobname="Ant_Area_wts" etc.

    I haven't put up the setup file for the third post yet. The main reason is that I had put a few lines in the main code to fix the area boundaries, and I need to find a way of making that possible within the general framework (or I could patch the inventory file).

    ReplyDelete
  22. Nick, I have a couple of better links to the ice core results that I linked previously with an article by the Idso's. The first is the paper of interest by Schneider and Steig and the second is the dissertation by Schneider when he was a student of Steig's at the University of Washington. The material in the dissertation indicates to me that it is Schneider who do a lot of the heavy lifting for a number of papers that followed, including S(09).

    http://www.pnas.org/content/105/34/12154.full

    http://www.ess.washington.edu/web/surface/Glacio
    logy/dissertations/schneider_dissertation.pdf

    I have just started going over your R code for the TempLS V2.1 algorithm used on the Antarctica data.

    ReplyDelete
  23. Thanks, Kenneth,
    As I mentioned in the latest post, I'll be travelling until early May (in the US, NY mostly). So I probably won't be able to read these till I get back. But I'll be very interested to hear how you get on with the code.

    ReplyDelete
  24. Nick, I am in the process of using your code to do some sensitivity testing. I'll report the results - unless someone else does it first.

    Your code was very well documented and would be breeze for anyone with good R skills to use straight out of the box. It took me a little longer to figure out how to modify the code and use it for MyJob outputs. I suppose R for Dummies might be in order for the likes of me, but it is satisfying to figure these things out for oneself.

    ReplyDelete
  25. Kenneth,
    Thanks for those upbeat remarks. I hope you find the input file structure helps for sensitivity testing. You should be able to expect that the user variables (u$...) won't be changed within the program, so you can set up multiple runs in which you just vary what you want to with each subsequent run.

    With the list structure, you can also set up a base state by writing, say, u0=u
    at the end of the first expression. Then whenever you want to return to the base state in a subsequent run, just set u=u0.

    ReplyDelete
  26. Nick, when you have some time could you send Eli the data for the weighted ground stations and the AVHRR graphs? elirabett2003@yahoo.com

    ReplyDelete
  27. Eli,
    Yes, I'll do that. I'm in the US now, and won't be home for a fortnight, and I don't have the data with me. I'll send when I get home.

    If you don't mind dabbling with R, you can get the plots with the latest version of TempLS. Since it is interactive, you can then get the data from the workspace or hack the script. There is a setup file - Antarctic.r for the first post, and Ant_Area_Wts for the second. The data required is in the zip file. I didn't include a setup file for this post, but it's fairly straightforward, though a bit longer.

    ReplyDelete
  28. Nick, after doing some preliminary sensitivity testing a few issues come into better focus.

    First of all I have to think that it might be best to determine whether the temporal AVHRR or ground station data is appropriate for calculating trends. I think one data source has to be more (much) correct than the other and doing a weighting to obtain a weighted average does not make much sense. That using all ground station versus all AVHRR data gives significantly different results is rather obvious. Another problem is that if we doubt the ground data in favor of AVHRR then the period pre 1982 becomes a major unknown and we then only have the period 1982-2006 (or to present) to talk about.

    I recall that O(10) uses 126 EOFs for measuring the spatial relationship of the ground stations to AVHRR data. (I think this is correct but perhaps NicL can confirm whether it is or not). In your program I was able to obtain results using 15 orthogs (20 orthogs did not work). The results with the only moderately higher number of orthogs, that your program allows, does change the trends by large amounts. Is it possible to modify the code so that I can look at higher numbers of orthogs? I noticed that with lower numbers of orthogs that I see what appears to be a bleeding of regional trends into adjacent regions.

    RyanO was able to show that by using the S(09) method that trends could be translated across regional boundaries. I was curious whether you can tell me how to look at the Antarctica without the Peninsula by modifying your R code. I was able to get to vector in your code that had, as I recall, [1:105] zeros and [106:1496] with all positive values. I assume I would have to have a comparable eofs data set that corresponded to no Peninsula data.

    ReplyDelete
  29. Kenneth,
    I agree that it's really a user decision about preference for AVHRR or ground, and that a relative weighting factor does not express any clear physical logic. The factor balances weights overall - before 1981 there are only ground stations, so they get 100%. As you say, if you don't believe that data, then you should not seek trends before 1981. The wtf rationalises how the two sets are
    combined across the 1981 break. And it does give a continuous merge factor so you can watch how the transition occurs.

    On numbers of EOFs, although Ryan used large numbers, he also used Tikhonov regularisation, so the kgnd parameter effectively restricted the degrees of freedom just as cut-off does here. I think there simply isn't enough information for more than say 10 EOF's to give sensible results, and when you see variation for more EOF's, that is noise from near-singularity. You can check this from the eigenvalue plots printed out, like those I showed here. When the condition number is getting high, you can improve the number by regularisation, but this isn't really giving you more information.

    You don't need to modify the code to eliminate the Peninsula. That's what the u$station_mask expression is for. You just include a logical requirement that u$mod>1. These descriptors are from the GHCN-style inventory, but for the Antarctic set I've modified their meaning. u$country is now 0 for ground stations and randomly 1-100 for AVHRR. u$mod is 1 for Penin, 2 for WA, three for Ross SEa, and 4 for East A (I'm writing from memory here - I don't have the file to hand).

    ReplyDelete
  30. Kenneth,
    memory lapse there - it's v$i$mod, not u$mod. Also v$i$country.

    ReplyDelete
  31. Nick, thanks much for showing me an easy way to exclude the Peninsula or just use the West Antarctic data with your algorithm.

    First let me say that I think there might be a problem with using the satellite temporal AVHRR data post 1981 with the temporal ground data pre-1982 to calculate trends. If we know that the AVHRR data gives warmer trends than the ground data and we then combine it with the pre-satellite ground data, do not we get an artificial positive trend over the entire 1957-2006 period due to an expected jump from the pre-1982 cooler ground data to the warmer satellite data after 1982?

    I am presenting a table of some of the sensitivity tests I have done for the 1957-2006 trends with all the Antarctic regions, with the Peninsula excluded and then for just the West Antarctic region (first link below). I only included those runs with wt =1 (temporally based on ground stations only) and wt=0 (temporally based on satellite data only). I do not see any good reason to use in-between weights.

    As you have noted previously, using a greater share of, or exclusively, the satellite data appears to stabilize the results obtained when varying the orthogs and even with the exclusion of the Peninsula and using just the WA data. I am a bit confused as I thought your algorithm worked with the ground data translated to the AVHRR grids through the eofs of spatial correlations and thus we are not looking at an effect from sparse data here. Increasing the number of orthogs with your algorithm does show, as you noted previously, some large trend variations with orthog numbers - when the wt=1, i.e. using ground data exclusively for temporal considerations. It makes me question even using 6 or 7 orthhogs with ground temporal data.

    That your algorithm bleeds trends across regional boundaries when the weighting uses temporal ground data exclusively can be seen in the results in the table linked below when the Peninsula is excluded or West Antarctica is looked at in isolation.

    I have included a graph from Steig's graduate student Schneider's dissertation that is linked below.

    The link to the dissertation as given before is here:

    http://www.ess.washington.edu/web/surface/Glaciology/dissertations/schneider_dissertation.pdf

    The graph, I believe represents the Antarctic trends without the Peninsula, but I am not certain. It shows that the trends are downward after 1982 for ice core and the satellite measurements which is not the case with what your algorithm shows for that period (I need to show that analysis later) and particularly so when you use the temporal AVHRR data exclusively for determining trends . It brings to mind that S(09) should probably have looked at the Antarctic data without the Peninsula and then shown the Peninsula with all the data that is available (as noted in that paper and in your blog here) for it as an aside. Also the trends from ice core data over the period from 1965 onward in Schneider's graph are also downward as might be expected from the other data in the dissertation link that shows the cyclical nature of the Antarctic temperature trends over a 200 year period.

    http://img576.imageshack.us/img576/9657/nsantpeneff.png

    http://img822.imageshack.us/img822/4721/nsantschneiderdes.png

    ReplyDelete
  32. Nick, thanks much for showing me an easy way to exclude the Peninsula or just use the West Antarctic data with your algorithm.

    First let me say that I think there might be a problem with using the satellite temporal AVHRR data post 1981 with the temporal ground data pre-1982 to calculate trends. If we know that the AVHRR data gives warmer trends than the ground data and we then combine it with the pre-satellite ground data, do not we get an artificial positive trend over the entire 1957-2006 period due to an expected jump from the pre-1982 cooler ground data to the warmer satellite data after 1982?

    I am presenting a table of some of the sensitivity tests I have done for the 1957-2006 trends with all the Antarctic regions, with the Peninsula excluded and then for just the West Antarctic region (first link below). I only included those runs with wt =1 (temporally based on ground stations only) and wt=0 (temporally based on satellite data only). I do not see any good reason to use in-between weights.

    As you have noted previously, using a greater share of, or exclusively, the satellite data appears to stabilize the results obtained when varying the orthogs and even with the exclusion of the Peninsula and using just the WA data. I am a bit confused as I thought your algorithm worked with the ground data translated to the AVHRR grids through the eofs of spatial correlations and thus we are not looking at an effect from sparse data here. Increasing the number of orthogs with your algorithm does show, as you noted previously, some large trend variations with orthog numbers when the wt=1. It makes me question even using 6 or 7 orthhogs.

    That your algorithm bleeds trends across regional boundaries when the weighting uses temporal ground data exclusively can be seen in the results in the table linked below when the Peninsula is exclude or West Antarctica is looked at in isolation.

    I have included a graph from Steig's graduate student Schneider's dissertation that is linked below.

    The link to the dissertation as given before is here:

    http://www.ess.washington.edu/web/surface/Glaciology/dissertations/schneider_dissertation.pdf

    The graph, I believe represents the Antarctic trends without the Peninsula, but I am not certain. It shows that the trends are downward after 1982 for ice core and the satellite measurements which is not the case with what your algorithm shows for that period (I need to show that analysis later) and particularly so when you use the temporal AVHRR data exclusively for determining trends . It brings to mind that S(09) should probably have looked at the Antarctic data without the Peninsula and then shown the Peninsula with all the data that is available (as noted in that paper and in your blog here) for it as an aside. Also the trends over the period from 1965 onward in Schneider's graph are also downward as might be expected from the other data in the dissertation link that shows the cyclical nature of the Antarctic temperature trends over a 200 year period.

    http://img576.imageshack.us/img576/9657/nsantpeneff.png

    http://img822.imageshack.us/img822/4721/nsantschneiderdes.png

    ReplyDelete