A Tale of Two Weather Stations

Down in the windswept wastes of the great Southern Ocean lies Base Orcadas, an Argentine naval outpost that boasts a weather station, and it’s a particularly important weather station because it’s the only one below latitude 60S with a long-term temperature record. The station was established by the Scottish National Antarctic Expedition in 1903 and transferred to Argentina in 1904 and it has been recording temperatures since April 1903. The station is located on a somewhat insecure-looking gravel spit on Laurie Island in the South Orkneys:

Figure 1: Base Orcadas Weather Station, South Orkney Islands: 60.7S 44.7W (Photo credit Noticias Ambientales)

The Orcadas record has another distinction: it’s one of the few records in the Southern Hemisphere where the homogeneity adjustments applied by GISS, NCDC and BEST generate a substantial cooling correction. Here we will look into whether this cooling correction is any more valid than the wholesale warming corrections, the adjustments applied elsewhere in the Hemisphere.

A note on data sources before proceeding. Here I use the GISS raw and adjusted GHCNv2 records from GISTEMP, the NCDC raw and adjusted GHCNv3.2 records from KNMI Climate Explorer and the BEST raw and adjusted records from BEST’s station data website. The three data sets give very similar results and are used interchangeably. However, BEST’s raw records contain more readings than those of GISS or NCDC and fill in gaps in the GISS and NCDC records. Where BEST got the extra data from is not stated.

Figure 2 shows the location of Orcadas relative to the two other reasonably complete temperature records in the area that go back to the early 1900s or before – Grytviken in South Georgia, located 900km to the northeast and Punta Arenas in Chile, located 1,800km to the northwest.

Figure 2: Location of Base Orcadas relative to Grytviken and Punta Arenas, Google Earth view.

The raw Orcadas record is a fairly good match to the raw records in the Antarctic Peninsula over the period of common readings after 1945, but not to Grytviken and Punta Arenas (Figure 3). Orcadas shows much larger short-term temperature swings than the other two records and also about 2C of net warming, compared to about a degree of net warming at Grytviken and effectively no warming at Punta Arenas:

Figure 3: Raw annual mean temperatures, Orcadas, Grytviken and Punta Arenas. Data NCDC.

It would be reasonable to conclude that these three records are located in areas with different temperature histories and should therefore not be homogenized, but they get homogenized anyway. The homogenization process adds about a degree of warming to Punta Arenas, subtracts about a degree of warming from Orcadas and leaves Grytviken substantially unchanged. The result is three reasonably congruous records, each of which shows about a degree of overall warming (Figure 4).

Figure 4: Adjusted annual mean temperatures, Orcadas, Grytviken and Punta Arenas. Data NCDC.

We will now take a closer look at the cooling corrections that the three data sets apply to Base Orcadas. We can quantify these corrections simply by subtracting the raw monthly data from the adjusted monthly data, and Figure 5 shows what we get when we do this. GISS identifies a 0.8C downward shift in the raw record (requiring a 0.8C warming “correction”) in September 1950, NCDC identifies a 0.7C downward shift in the same month and BEST, which adds some kind of a residual seasonal adjustment during the homogenization process – hence the wiggles – identifies a downward shift of 0.65C in December 1949:

Figure 5: Cooling corrections applied to Orcadas raw record

How do the homogenization algorithms come up with these corrections? BEST gives the most detail on procedures so I will use BEST as the example. First BEST presents a plot of the raw Orcadas record, which shows gradual warming since about 1930 and no obvious shifts:

Figure 6: BEST Orcadas raw record, monthly means

Then BEST subtracts the Orcadas raw record from a“regional temperature expectation” or “regional average” series constructed using records from up to 2,000km away, and from the result it deduces the existence of an artificial upward shift of about 0.7C in the Orcadas record in December 1949. Note that the shift is classified as an “empirical break”, i.e there is no record of a coincident station move or time of observation change:

Figure 7: Difference between BEST raw Orcadas record and BEST “regional average”  for Orcadas

BEST then adjusts the shift out, and to confirm it did it right it presents a plot comparing the adjusted Orcadas record with its regional temperature expectation series. Not suprisingly, the two match:

Figure 8: BEST adjusted Orcadas record versus BEST regional average

However, this approach involves more than a little circular logic, and BEST also assumes that records from anywhere within a 12.6 million square kilometer area centered on Orcadas can be used to define the “regional expectation” in the South Orkneys, which is a stretch. Consequently we are left with the possibility that the shift is simply an artifact of the homogenization algorithms. How do we find out whether it is?

Well, we could do it fairly easily if there happened to be a station near Orcadas with a record covering the period around 1950. And as it happens, there is:

Figure 9: Signy Island Weather Station, South Orkney Islands: 60.7S 45.6W

Located according to BEST a mere 47.33km west of Orcadas is the Signy Island Research station, located on what appears to be somewhat firma terra than Orcadas. Signy is run by the British Antarctic Survey and is on the site of an old South Orkneys whaling station. It has been in operation since January 1947 and has a temperature record running from then to December 1995, when the station was put on a summer-occupancy schedule.

So how do the Signy and Orcadas records compare? Figure 10 plots BEST’s versions of the two raw records over the period of common readings. The records track each other remarkably well for most of the time considering the highly variable temperatures, which jump around from year to year as much here as anywhere else on Earth, particularly in the winter. The generally high level of congruity and the lack of any large displacements in the difference plot suggests that both records are measuring real temperature variations and are not significantly affected by artificial distortions:

Figure 10: Orcadas versus Signy raw records, BEST monthly data

The only feature that looks as if it might complicate matters is the “bust” in 1950, but this turns out to be a simple quality control failure. As shown in Figure 11 the Orcadas and Signy raw records are displaced by a month relative to each other during most of 1950 in all three data sets, and GISS’s, NCDC’s and BEST’s QC procedures didn’t pick up on it.

Figure 11: The one-month shift between the Orcadas and Signy raw records during 1950, BEST monthly data

I corrected the displacement by deleting the implausible-looking December 1950 value in the Signy record and moving the Signy values after February 1950 forward a month (I may have moved the wrong record but as a practical matter it makes no difference which one is shifted). The corrected comparison is shown in Figure 12:

Figure 12: Orcadas and Signy raw records after correction for shift

Subtracting Signy from Orcadas now yields the results shown in Figure 13. If the Orcadas record contains a ~0.7C artificial shift in or around 1950 we should see it clearly in the difference plot at the bottom, but we don’t.

Figure 13: Orcadas versus Signy raw records after correction for one-month shift.

But we see a clear displacement in late 1950 when we compare the raw Signy record with any of the three homogeneity-adjusted Orcadas records (Figure 14 compares it with the GISS version). There can be no clearer proof that the shift is manufactured by the adjustments:

Figure 14: Orcadas adjusted record versus Signy raw record, BEST data

And because the raw Signy record does not receive any adjustments what we see in Figure 14 is the end product of the homogenization process. Instead of turning a distorted raw record into an undistorted adjusted record homogenization has turned an undistorted raw record into a distorted adjusted record. And the adjusted record is supposedly the one that’s correct.

This entry was posted in Climate change and tagged , , , , . Bookmark the permalink.

17 Responses to A Tale of Two Weather Stations

  1. Willem Post says:


    Great sleuthing.

    Are there other examples with such “adjustments”. What if enough adjustments took place to establish a trend?

    The above 12.6 million sq km compares as follows:

    “The 48 states and D.C. together have an area of 3,119,884.69 square miles (8,080,464.25 km²). Of this, 2,959,064.44 sq mi (7,663,941.71 km²) is land, comprising 83.65% of U.S. land area.

    There may be areas in the US that warmed less than others.

    Vermont, April 6, and it is STILL freezing at night.

  2. Dave Rutledge says:

    Hi Roger,

    A great addition to your series. From studying the US records and from the early debate on Chinese UHI, I was prepared for the possibility that the land temperature indices were a mess. But your posts and Euan’s have helped me understand the problem better.


    • Thanks Dave, but I don’t know whether it qualifies as great. It’s just another step on the road towards developing a better understanding of how the “corrections” applied by GISS, BEST et al generate the results they do. So far I’ve reached the following conclusions:

      1. GISS, BEST at al. are not deliberately trying to falsify the surface air temperature record.

      2. The distortions introduced by the corrections are a result of flawed homogenization algorithms that sausage-machine thousands of raw records apparently without anyone bothering to look at these records before pressing the “run” key.

      3. The damage is most likely done during the iteration/convergence processes that generate the “regional expectation” or “reference” series which the raw records are adjusted to match.

      4. The question is why the flaws, which are pretty obvious, have not been recognized, or at least not owned up to, by GISS, BEST et al.

      I hope to be able to tighten these conclusions up with more study but can’t guarantee that I will be able to add much.

  3. Euan Mearns says:

    Hi Roger

    This is a message from Euan. He is ski-ing for a couple of days in the Scottish mountains and has no access to wifi. He has prepared a post which should be ready for Wednesday.

    Kathryn Mearns

    • Thanks Kathryn. Happy internet-free skiing to Euan and tell him not to break anything.

      • Kathryn Mearns says:

        Hi Roger

        Euan has just called. There is currently no functioning wifi on the west coast of Scotland. Unfortunately, no post tomorrow.


  4. ristvan says:

    Roger, wrt your points 3 and 4, my own assessment (essay When Data Isn’t) has identified 3 generic homogenization logic flaws. Perhaps most important is regional expectations. BEST 166900 and Aus BOM Rutherglen provide concrete examples, as does Euan’s post here. The regions are too large, and can contain disparate ‘microclimates’ no different than French wine terroire. Second is UHI, which is really a misnomer for land use change indicated by population density. Two separate California studies are discussed in the essay. The net result for the US is shown by the surfacestations.org project and resulting paper. This means ‘UHI polluted’ stations can form the majority of a regional expectation, which then guarantees the ‘pollution’ is homogenized into unpolluted stations. Examining the GISS homogenization of all the surfacestations.org CRN1 CONUS stations proves this point (essay written, not yet posted anywhere, although Curry has a copy). Third is Menne stitching, used to resplice records in most homogenization algorithms. Zhang et. al. Illustrated (in China) and explained the warm biased logical flaw in their 2014 Theor. Appl. Climatol. paper.
    I suspect simple confirmation bias is why these things have not been more closely scrutinized by the climate establishment. The results matched their expectations and ‘felt right’.

    BTW, I find the whole homogenization thing a tempest in a teapot compared to model falsifications like the ‘pause’, the missing tropical troposphere hot spot, the lack of accelerating sea level rise, and the lower observational energy budget TCR and ECS meaning there is no problem to worry about.

    • Ristvan: Do you have links to the studies you cite? I’d like to take a look at them.

      I tend to agree that confirmation bias is a large part of the reason why these things have not been more closely scrutinized.

      • ristvan says:

        See the extensive footnotes to essay When Data Isn’t in ebook Blowing Smoke: essays on energy and climate. Available cheap iBooks, Amazon Kindle. Nook, Kobo, …Get me somehow your coordinates and I will send the unpublished CONUS surfacestations.org CRN1 analysis for you to use as you see fit with attribution. Judith Curry would be a good intermediate, as she has guest posted partials of many of my essays as well as one recently from Euan, and obviously has both our private email coordinates.

    • manicbeancounter says:

      Your last comment about homogenization being a “tempest in a teapot” may be true if you look at the adjustment to the global average. However, as you yourself have noted, and as noted above, the homogenization adjustments literally “make homogeneous” or smooth out what are clearly local variations in the data. It is more than confirmation bias. Rather, it is imposing on the data the belief that what such micro climates exhibit is data noise, probably due to some unknown measurement bias. Thus the homogenized data smooths away the true diversity and complexity of the real data. As such homogenized data ends up as part imposed opinion, bringing the data closer to theory than an unbiased data set. So when the question is asked of how much warming is human caused, the answer can change from a trivial amount, to more than 100%.

    • ristvan says:

      G3, this is a fascinating new Australian analysis. As another data point, the Steriou et. al. presentation at European AGU 2012 estimated 1/3 rather than 2/3 (global) based on a GHCN sample that included all stations with records back to 1900 (except in the US, which would otherwise have been over sampled). See essay When Data Isn’t plus footnote 13, or go to http://www.itau.ntua.gr/en/docinfo/1212. for a copy of the paper as presented.

    • Here are the results of my initial analysis of the BOM “corrections” to Australian raw temperature records:

      I put it together eleven years ago in 2004.

  5. Pingback: GWPF Launches Enquiry into Integrity of Temperature Data | Energy Matters

Comments are closed.