The Hunt For Global Warming: Southern Hemisphere Summary

In recent months I’ve had a series of posts looking at the temperature histories of a number of land areas in the Southern Hemisphere [1, 2, 3, 4, 5]. This was in response to a post by Roger Andrews where he presented an analysis of about 300 climate stations from the Southern Hemisphere that, when combined, showed substantially less warming than the reconstructions presented by various groups (BEST, GISS, HadCRUT) [6]. I found this to be both intriguing and important and wanted to see if I could replicate Roger’s result.

Figure 1 Note that this chart has an expanded Y-axis scale and the grid lines are at 0.1˚C intervals. A regression through all the data using station average as the base indicates warming of +0.18˚C per century, i.e. close to zero. The black trend lines are parallel to the regression and show there are rising tops and bottoms in an overall slowly rising trend. The alternative view is a flat trend from 1880 to the mid-1970s with a step change to warmer temperatures across the mid-1970s cold snap.

  • The average time-temperature anomaly series for 174 climate stations from New Zealand, Central Australia [1], Southern Africa [2,3], Patagonia [4], Paraguay and Antarctica [5] are presented in Figure 1. A simple regression through the data with no weighting indicates warming of +0.23˚C since 1880 equivalent to +0.18˚C per century. This is substantially less than S hemisphere land temperature reconstructions reported by BEST, GISS and HadCRUT.
  • Comparing with Roger Andrews’ reconstruction the difference is less than 0.1˚C. I have managed to replicate Roger’s result.
  • My default method is to use each station’s average temperature as a base for calculating anomalies. All anomalies have also been calculated using a fixed base period of 1963 to 1992. Doing so makes no material difference. Warming is reduced to +0.21˚C since 1880 using the fixed base.
  • Area weighting the results produces a trend of +0.19˚C warming since 1880. Area weighting lends substantial weight to Antarctica (only 14 records) where the main data series begin in 1954 and possible methodological problems are identified with lending 55% weight to only 14 stations with records that begin in the mid point of the time series.
  • The mid-1970s stand out as an anomalous cool period seen in records throughout Central Australia and Southern Africa. It was also cool, although not anomalously so, in S America, Antarctica and New Zealand at this time.
  • The structure of the data is one of cyclical warming to circa 1914 followed by cyclical cooling to the mid-1970s followed by cyclical warming to the present day. There is no evidence for warming linked to the 1998 el nino in these data. Nor is there evidence for a pause in warming in the Southern Hemisphere since 1998.
  • The data presented here are not intended to provide full cover of the Southern Hemisphere. Coverage will be extended at some point. But there is probably sufficient cover to be representative of the southern hemisphere land and surrounding marine conditions. There is scant evidence for significant warming in these data suggesting global warming is confined mainly to the Northern Hemisphere.


GHCN V2 data as downloaded from the NASA GISS web interface are used throughout. These V2 data have been subjected to minor adjustments by GHCN, which is undesirable, but the level of adjustment is substantially less than the V3 homogenised data.

The areas of study have been selected to as far as possible avoid large numbers of people and reworking of the land surface by Man although this has not always been possible, for example in Southern Africa. Urban records present a dilemma. The most valuable records are long and continuous and these are often from urban areas. Urban records that show clear signs of urban warming were not used. It was only in Southern Africa that 7 urban records were rejected on this basis.

In many cases I allowed the NASA GISS station selector map to choose the stations for me. For example placing Alice Springs at the centre of a NASA GISS search returns 30 records in a 1000 km radius around Alice Springs. This was done in New Zealand, Patagonia and Paraguay. However, in Southern Africa the GISS interface returns very large numbers of short rural records and I therefore clicked around on the map searching for the longer continuous records. Similarly in Antarctica I clicked around on the map looking for the longer and continuous records. In addition to the continental areas, the islands of South Orkney (Base Orcadas), South Georgia (Grytviken) and Kerguelen are included.

Most of the individual areas have been presented before as individual posts  [1, 2, 3, 4, 5]. New Zealand and Paraguay have not and these data are presented as appendices to this post.

Figure 2 S Hemisphere map showing the distribution of areas sampled. These have in general been chosen to avoid large centres of human population and prosperity. I will return to sample more areas and it will be interesting to see to what extent population density and / or latitude impacts the results. 

The time series distribution of stations is shown in Figure 3. While 174 stations have been examined, the maximum operational at any one time was 146 in the mid 1970s. The minimum number was 7 in the early 1880s and this may impart some small sample bias to the earliest part of the composite record. From 1889 there were at least 12 operational stations and small sample bias diminishes thereafter as increasing numbers of stations come on line.

It is assumed that the commissioning of stations was a random process and that station number growth to the mid 1970s should not impart any bias to the composite data with one exception. Antarctica, that represents 55% of the total land area (of the sampled land area), only came on line with 14 stations around 1956. If those data are area weighted this results in a major data structure discontinuity.

A source of possible concern is the mass program of station closures that took place in 1990/91. We go from 118 to 72 stations in 2 years. If this process was non-random then it has potential to bias results from that time onwards.

Figure 3 The distribution of operational stations from the group of 174 selected stations.


With discontinuous time-temperature records it is essential to convert station temperatures to anomalies to minimise the effect of the discontinuous series. The conventional way to do this is to use the mean temperature for a base period, for example 1951 to 1980 favoured by GISS and others, and to deduct the mean temperature from that base period from the whole time-temperature series to produce a time-anomaly (dT) series. This is undoubtedly one approach. But what if a station record has no data for the base period?

Approaching this as a novice some months ago I decided the best and simplest way to do this was to use the average temperature for a station as the base. This uses all the data in the station normalisation procedure which I instinctively feel is correct. One consequence of this is to have a non-uniform base time and it was anticipated that this procedure may suppress differences while using a fixed base period may amplify them. There is no absolutely correct way around this. When my Central Australian post appeared on Climate Etc some weeks ago there was a disproportionate amount of interest in “flawed normalisation” as a scapegoat for what the data really showed.

And so for the S Hemisphere I have re-calculated all records using a fixed base period of 1963 to 1992 (at the recommendation of Roger Andrews) which is a time period that catches most but not all records. Where a record is not represented in 1963-1992 I have used the station average instead. As we shall see none of this has any significant relevance to the substance of the debate.

Figure 4 Using a 1963 to 1992 base period for normalisation has no material effect on the results. Somewhat surprisingly the slope is slightly reduced compared with the station average base (Figure 1).

Figure 4 shows the stack of 174 stations replotted against the 1963 to 1992 base period. You will be hard pressed to see any difference between this and Figure 1 that uses the station average base. A regression through the 1963-92 base data produces a +0.21˚C since 1880, slightly but not significantly less than the station base method.


The main results have by now already been presented. These large areas of the southern hemisphere land mass show little significant warming since 1880. To place this in context the results are re-plotted at a more conventional scale in Figure 5. This shows that from 1880 to 1973 (almost 100 years) the trend was effectively flat. In the mid 1970s, centred on 1976, something strange happened to Southern Hemisphere climate. A marked cool period, accompanied by higher rainfall, gave way to an era of marginally higher temperatures, perhaps 0.2˚C warmer than the previous era. This is what climatologists should be seeking to explain. I do not believe it has anything to do with mankind’s activities.

The cool feature centred on 1976 is present in most Central Australian and Southern African records and is very real. It is probably over – represented in this data set since I have 79 records from these two areas (45% of the total records for 31% of the land area). But it was also cool, although not anomalously so, in S America, Antarctica and New Zealand at this time. Roger Andrews, Crutem4 and BEST all pick out this anomaly (see below) but GISS Temp does not. Evidently GHCN V3 homogenisation has homogenised this feature out of existence.

Figure 5 This chart plotted at different Y-axis scale provides an alternative perspective.

I have marked a few other events on Figure 5. Krakatoa erupted in August 1883. 1884 was a little cool in the S hemisphere, but not uncommonly so. 1891 was as cool. The mass station closures of 1990/91 do not really appear to have affected the structure of the time-temperature profile. The 1998 El Niño temperature spike that all climate watchers are familiar with also appears to be absent in these S hemisphere records although present in GISS Temp and Crutem4. As discussed below, the presence of the 1998 El Niño spike in Crutem4 looks particularly suspicious since it is quite clearly absent in the 174 records analysed here.

Many workers and some commentators like to see a smoothed moving average. There is certainly a place for this with certain kinds of data. The S hemisphere land data is not noisy and does not really require smoothing but a centred 5Y moving average is shown in Figure 6 nonetheless.

Figure 6 A 5Y centred moving average reveals the general structure of the data. A regression from 1880 to 1980 actually shows a gently cooling trend. If CO2 radiatively forces temperatures then it failed to do so in these parts of the S hemisphere during that 100 year period. This chart does show a little warming since 1980 but don’t be deceived by the scale. Recent peaks are +0.3˚C compared with +0.1˚C back in 1914. Recent warming is real but trivial.

Area Weighting

The professional reconstructions build a global grid and seek to allocate a temperature to most or all cells within that grid in some cases regardless of whether or not there is a temperature record for it. The temperature history for each grid cell is weighted according to the grid cell area. This appears to be a sound methodology until one actually attempts to do this, which then makes one aware of the limitations.

With the current set of records, Antarctica exemplifies this problem. For the main continental area the main record series begins in the mid-1950s. We quite simply do not have a temperature record for the main continent before then. There are longer records from the Antarctic peninsula and the islands to the NE (South Orkney and South Georgia) but this area is quite clearly in a totally different climatic regime [5] and should absolutely not be used to model or project temperatures southwards to The Continent.

Figure 7 Areas and weights used to calculate the area weighted temperature-time series (Figure 8). Some of the areas are approximations. Prior to the mid 1950s the reconstruction shown in Figure 8 does not include Antarctica. Area weighting lends a large amount of significance to the 14 Antarctic records from the mid 1950s on.

The philosophy I have followed in selecting “good” records from a particular area is to identify records that exhibit similar trends or features. For example, records from Central Australia and Southern Africa that show the mid 1970s cooling bear a hall mark of good data quality. In working across a geographic area one develops a sense that all records are recording more or less the same temperature history, absolute temperatures being controlled by latitude and altitude. Each of the areas shown in Figure 2 can be regarded as having congruous records with the exception of South America where there is a higher degree of interwoven variability.

The area of each group shown in Figure 2 has been estimated and the area used to weight the mean time-temperature series for that group (Figure 7). The weight distribution varies as the time series come and go. For example the weights of Central Australia and Southern Africa are halved when data from Antarctica arrive in the mid 1950s (Figure 7). The outcome of this exercise is shown in Figure 8. The gradient is little changed but the distribution is more variable because of the different way the calculation is done and the weight given to the more variable Antarctic records.

Figure 8 The area weighted chart is not analogous to Figure 1. Figure 1 is the average of the stack of 174 records. This chart is the sum of weighted averages for 7 areas + the three islands (that have tiny weight). Warming of +0.15˚C per century is little changed from Figure 1. Note that prior to 1956 there was only single station patchy records for Antarctica and since these records are lent 55% weight the Antarctic series begins in 1956 when there were 5 operational stations.

Comparison With Other Reconstructions

One of the main objectives of this project is to try and get to the bottom of the warming seen in time-temperature series for the Southern Hemisphere produced by large well funded groups like NASA GISS, UK HadCru and BEST. I have tortured the data for the 174 stations I selected every which way I know and cannot squeeze more than +0.18˚C per century out of them. I have been told repeatedly that I am wasting my time. So many EXPERTS working this data all come up with the same conclusions. Well not quite. There  is about as much difference between EM and CRUTEM4 as there is between CRUTEM4 and BEST/GISS Temp.

Figure 9 Roger Andrews’ reconstruction had about 300 records and more complete cover of the Southern Hemisphere. He is using a different base period that accounts for the gross offset between the two series (this applies to all the comparisons). There is a  high degree of congruity, i.e. the series are to large extent going up and down in unison.

Figure 10 Deducting EM from RA produces a fairly flat trend with a gradient much less than 0.1˚C per century. The main difference between myself and Roger probably lies in the greater geographic cover in his data series.

Figure 11 There appears to be an even higher degree of congruity between EM and CRUTEM4 (Crutem4 data kindly supplied by Roger Andrews) suggesting that similar data lies beneath these to reconstructions. CRUTEM4 for example picks out the mid-1970s cooling. One exception is the 1998 El Niño. CRUTEM4 has a large temperature spike in 1998 that is absent from the 174 records I examined. CRUTEM4 begins below and ends above the EM  series giving rise to a warming trend that is evidently absent in the “raw” records.

Figure 12 Deducting EM from CRUTEM4 produces a warming differential of about +0.5˚C between the two. This is the closest match I have to any of the “official” reconstructions. Trying to track down the origin of this warming that appears to be absent in the “raw” GHCN V2 records is one of the primary objectives of this project.

Figure 13 There is also a high degree of congruity between BEST and EM but the BEST series begins way lower and ends way higher (once again thanks to Roger Andrews for the annualised BEST land S hemisphere series). The high degree of congruity suggests that both they and I must be doing something right. But the origin of the gradient in the BEST data that seems absent in the “raw” records Roger and I have analysed remains a mystery. I am hopeful that Judy Curry may run this post and that Stephen Mosher who blogs there may begin to provide some form of explanation.

Figure 14 Deducting EM from BEST produces a warming differential of about +0.9˚C between the two.

Figure 15 GISS Temp probably gives the poorest match to the EM data series. This time the GISS Temp data comes from the NASA website from a link I believe Gavin pointed me to. As with the other “official” series GISS begins below EM and ends higher producing a temperature gradient that seems absent in the raw records that both Roger and I selected.

Figure 16 Deducting EM from GISS Temp produces a warming differential of about +0.9˚C between the two.

I believe that myself and Roger Andrews have sufficient data to provide a representative picture of Southern Hemisphere land temperature history and our data shows a warming trend of between +0.2 to +0.3˚C since 1880, significantly below the “official” reconstructions. The most perplexing is BEST who claim to be a group of sceptics that set out to test the veracity of GISS and HadCrut and yet have managed to generate warming from what are effectively flat records. The objective of this project is to try and discover where any errors lie. While non-representative cover may be a small part of that story I am not prepared to accept this as the whole explanation. I believe the answer may lie in a combination of the following:

  1. Homogenisation of data (adjustments of raw records) that is known to add +0.3 to +0.5˚C to the GHCN V3 series even although the main man-made artefact is UHI that should result in the application of a net cooling correction. Is the GHCN v3 homogenisation residual evenly spread across the globe? Or is it over-represented in the S Hemisphere to mask the inconvenience of little warming across half of Earth?
  2. The inclusion of UHI warmed urban records
  3. The adjustment of records to a regional expectation (BEST). Does this for example adjust S hemisphere records to a N hemisphere expectation? And are UHI warmed urban records included in that expectation? Has the CRUTEM4 1998 spike been imported from the N as an expectation?
  4. The projection of temperatures into grid cells where there are no records.

I believe the most reliable way  to get an accurate picture of global time-temperature is a simple analysis of what are viewed as the most reliable raw records. A degree of processing to create anomalies is required and area weighting may help ensure representative cover. Any processing beyond that opens the door to the introduction of biases and since further processing is not necessary it should be avoided.

[1] Temperature Adjustments in Australia
[2] The Hunt For Global Warming: Southern Africa
[3] The Hunt For Global Warming: Southern Africa Part 2
[4] The Hunt for Global Warming: South America
[5] The Hunt For Global Warming: Antarctica
[6] Homogenizing the World

[Note added 14th April: One of the comparisons I forgot to do was with the satellite data. The chart shows UAH, lower troposphere, land only, hopefully annual mean. I think this is pretty good and tends to conform the veracity of the satellite record. NOAA NCDC]

Appendix A New Zealand

Figure A1 15 climate stations in New Zealand selected by the NASA GISS interface.

Figure A2 T spaghetti for New Zealand. There are 5 long records and visual inspection shows no clear gradient.

Figure A3 dT spaghetti for New Zealand. There is a high degree of congruity between the records, especially post-1905.

Figure A4 T anomalies calculated relative to station base mean and 1963 to 1992 base period. There is clearly little difference between the two although the 63 to 92 base has a slightly steeper gradient. While a regression shows a slight gradient the overall trend is flat since it lacks rising tops and bottoms. 

Appendix B Paraguay and surrounding areas

Figure B1 34 climate stations in Paraguay and surrounding countries selected by the NASA GISS interface placing Puerto Casado at the centre. Many of these have a somewhat chaotic record and it was difficult to select good from bad and so the simple solution was to use them all.

Figure B2 T spaghetti for Paraguay and surroundings. There is a degree of congruity since many of the spikes go up and down together. But overall there is a higher degree of variability in these records compared with other areas. See for example Figure A2.

Figure B3 dT spaghetti for Paraguay and surroundings. A spike down can be seen in 1955 but otherwise the dT trends appear quite chaotic lacking any clear gradient.

Figure B4 T anomalies calculated relative to station mean base. Despite the more variable nature of the individual records this all seems to come out in the wash and the average of the stack is fairly typical for the Southern Hemisphere with a small positive gradient and gently rising tops and bottoms.


Roger Andrews provided initial inspiration and back up data. My son Neil Mearns compiled the New Zealand and Paraguay records and charts.

Note added 23rd April

Since making this post I have conducted a number of methodology checks that can be found at the link below:

Averaging Temperature Averages

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

32 Responses to The Hunt For Global Warming: Southern Hemisphere Summary

  1. This is a convincing presentation. You have shown the importance of determining rates of warming specific to hemispheres.

    If I understand correctly, the estimated global rate since the onset of the industrial revolution around 1850 until the present is about 0.5 K per century, more than the SH rate, which is what we would expect because of the imbalance in the distribution of land and sea in the two hemispheres.

    The increase in the southern hemisphere since 1880 seems to have been about 0.23 K per century.

    I am now looking at the global increase between 1600 and 1800 based on non-tree-ring proxies (Loehle, 2007, 2008). Global warming of 0.3 K per century between 1600 and 1800 seems to have been more rapid than SH warming since 1880. but less rapid than the global increase since 1850. However, the study by Loehle has mainly NH proxies.

    What would be interesting is to compare NH after 1880 with NH before 1880 and SH before and after the same dates.



  2. Euan,

    it would be surprising to me if you got more significant warming than that observed by NASA or others… 😉


  3. Roger Andrews says:

    I have previously estimated SH surface air temperatures a number of different ways. Here are four of them:

    First difference – using 369 raw records of variable quality
    Verified – using 269 raw records and record segments verified as congruous, adjusted to a common 1950-1990 baseline and area weighted relative to “climate zones”.
    First difference verified only – using the 269 congruous raw records and record segments
    Latitude weighted – all 369 raw records weighted by the cosine of latitude.

    The four series are plotted in the graph below. Temperatures are zeroed relative to 1996-2005 means so that any differences become apparent in the early rather than the later years.

    They all give essentially the same results after about 1900 but rapidly break apart before then. What this shows is that we don’t have enough good records to construct a reliable SH air temperature time series before about 1900. (The early parts of the BEST and CRUTEM4 SH air temperature series, which go back to 1855 and 1850 respectively, are pure speculation.) And if we don’t have enough good records to construct a SH air temperature series before 1900 then we don’t have enough to construct a global air temperature series before 1900 either.

    Now I superimpose Euan’s series on my four series.

    Euan’s approach was quite different to mine. He constructed his series by expressing temperatures as anomalies relative to the mean of the entire record regardless of the reading interval and by taking a simple average of the results. He also used only 174 records compared to my 269-369. Yet our series, while not exactly the same, are very similar. We can even explain why Euan shows a little less warming than me It’s because his 174 stations, which he selected to avoid populated areas as much as possible, weight the average towards the southern part of the SH, which has warmed less than the northern part.

    The take-home message here is that within reasonable limits it doesn’t much matter which raw SH records we use and how we average them together. We get pretty much the same results whatever we do. The data have spoken.

    Now contrast our results with the SH air temperature series published by GISS, NOAA/NCDC, CRU and BEST (below). They all show more warming than Euan and me. Why? Because they apply spurious warming “adjustments”; there’s no other way they can do it. And not only that, they can’t even agree on how large the adjustments should be (NOAA shows almost twice as much warming as GISS). Which brings up another problem. Which of these august scientific institutions is right? Do we even have a reliable published SH surface air temperature series? If we don’t we don’t have a reliable global one either.

    • Euan Mearns says:

      Roger, interesting comment, I like the second chart. The main point is that the majority of these S hemisphere temperature records are “flat” and so no matter how you slice and dice them you get a “flat” outcome. Unless of course you work at Berkeley. I’ll be sending this to Richard Muller tomorrow inviting him to comment. And to Phil Jones.

      I noticed that HadCRUT area weight the Antarctic stations in a single grid cell around the station leaving about 14 M km^2 of Antarctica un-represented even though the katabatic winds blow from centre to edge pretty well year round.

      • Euan: The CRUTEM4 series in the Antarctic is a little smoother than the GISS, NCDC and BEST series but is otherwise similar, and all of them more or less match the raw records. So I don’t think the single grid cell approach makes much difference. (Note that “Had” isn’t involved here – land series are all from “CRU”).

        Best of luck with Muller and Jones.

    • manicbeancounter says:

      You are puzzled as to why simple averaging obtains far less warming than the scientific institutions. I think you were part way there when you looked at the BEST South American data last month.
      Here you looked at the BEST methodology for temperature adjustments. They build up an expected regional anomaly, then (through a highly weighted iterative procedure) adjust data according to that weighted average. What is does not do is remove most of the year-on-year fluctuations, which Euan has noted are consistent across much of the Southern Hemisphere. Also (as you have noted in comments) BEST seems to take account where there is a clear UHI bias. At a superficial level the fact that BEST in homogenizing the “raw data” increases the warming trend over 125 years four-fold does not make sense. The homogenization process seems pretty rigorous. I throw out a possible explanation that needs to be tested against the data. The BEST homogenisation procedure assumes a tight bell-shaped distribution of individual weather stations from the regional anomaly. In the real world there are micro climates, so the distribution might be approximately bell-shaped, but flatter with smaller peaks in the tails. The raw data is also skewed by UHI. Probably a majority of weather stations will have some measure of UHI. The temperature stations with extreme UHI will be adjusted, but not to the full extent, as the regional expected average will be part influenced by moderate UHI. On this suppose there are two weather stations in an area with a micro climate. The first accurately exhibits a drop in average temperature of a degree Celsius from 1980 to 2000, the other a degree rise. The true (but unknown) regional rise is 0.2C and the expected average is 0.4C. I would suggest that the adjustment for the first data set would be >1.2, whilst the adjustment for the first data set would be < 0.4. A 1.4C variance from the expected would be moved much closer to the expected than a 0.6C variance due to the aggressive weight procedure.
      My hypothesis might be testable, at least for extreme cases, or where we have independent estimates of measurement biases such as for Iceland.

      • Roger Andrews says:

        Manic: Does Figure 10 of the Worst of BEST post shed any light on your theory?

        • manicbeancounter says:

          Figure 10 sort of shows what I am getting at. You show a scatter of raw record trend against BEST adjustment. What I was thinking about was putting the raw trends and adjusted trends into bands of say 1/4 and showing the counts on a histogram. It will give a bit more detail to you observation that all the negative trends were eliminated. The BEST adjusted will be strongly clustered around a degree, but it should also show that the raw trends are predominantly below a degree.
          If you send me (via Euan) data for figure 10, I could run up a graph.

      • Euan Mearns says:

        Figure 10:  XY plot of raw record trend line gradients for 86 South American stations versus adjustments applied by BEST

        Here is Roger’s Figure 10. How did you make this Roger? And should I send it to Mosh?

        The difference between the S Hemisphere warming gradients observed by Roger and I and those from other institutions is substantial and cannot be explained by trivial errors. My leading contenders to explain difference with BEST:

        1) The GHCN V2 data are not fit for purpose
        2) BEST import temperature trend data from the N Hemisphere
        3) Homogenisation adds bias. Perhaps homogenisation results in zero bias globally but is skewed between the hemispheres?

        Figure 10 is intriguing. Do you have same for Southern Africa and Australia?

        • Roger Andrews says:

          Hi Euan: I made this plot by fitting linear regression lines to the raw and adjusted BEST records. It’s a simple but laborious process. If your spreadsheet will calculate gradients through discontinuous data (mine won’t) you can get the numbers by pressing a button. If not you have to measure the gradient off a regression line on a graph. And of course you have to download the BEST records before you start.

          I have some numbers for S. Africa somewhere but they don’t use the BEST data. The S America graph does.

          Things a little busy here at the moment. Back with more stuff ASAP.

          • Euan Mearns says:

            I just had a look at Curtain Spring. A quicker way of doing this may be to use the warming trends / century reported by BEST – these are close to the regressions I did for Curtain. Also, in Australia most records appear to have been substantially cooled. I still can’t get my head around folks taking this raw data and applying such major corrections to it before using it. And I can’t as yet get my head around BEST cooling the raw records in Australia and still getting more warming out of them.

          • Roger Andrews says:

            Yeah, my mind is elsewhere. That’s what I did.

        • Roger Andrews says:

          Here’s a numerical example of how I think the BEST SH warming bias may originate.

          About two-thirds of the raw S American records show warming and about one-third cooling, and because of the enormous 2,000km search radius BEST averages hundreds of these records together to calculate a time series in a single grid block. To simplify matters I assume here that we have only three records, two showing a warming gradient of 0.9C and one showing a cooling gradient of 0.9C (doesn’t matter what the time interval is). For further simplification I assume that the kriging weights for the three stations are the same.

          In its first run the BEST algorithm averages the three records. It calculates an average warming gradient of (0.9 + 0.9 – 0.9)/3 = 0.9/3 = 0.3C – the initial “temperature expectation” gradient – and a standard deviation of 1.04C.

          In accordance with instructions received the algorithm now starts to de-weight the less reliable records. It finds that it has two records with 0.9C of warming, 0.6C higher than the temperature expectation, and one with 0.9C of cooling, 1.2C lower than the expectation. It therefore identifies the cooling record as being less reliable than the warming records and deweights it relative to the warming records – we will assume by a third, so that it contributes only 0.6C of cooling instead of 0.9C.

          Now it starts iterating, seeking convergence. It measures a warming gradient of (0.9 + 0.9 – 0.6)/3 = 0.4C, up a tenth, but the deweighted record is still farther away from the expectation than the other two records (1.0C versus 0.5C). But the standard deviation is also lower – 0.87C instead of 1.04C, showing that things are beginning to converge.

          So the algorithm deweights the cooling record some more, let’s say by two-thirds, so that it now contributes only minus 0.3C. Now the temperature expectation becomes (0.9 + 0.9 – 0.3)/3 = 0.5C, but the cooling record is still farther away from the expectation than the two warming records (0.8C versus 0.4C) and the standard deviation has dropped to 0.69C, showing that the records are still converging.

          So the algorithm continues to deweight the cooling record, and the more it deweights it the more convergence it gets. But exact convergence isn’t reached until the cooling record is deweighted to zero, which leaves us with a temperature expectation of 0.9C of warming.

          Now I have to assume that the BEST algorithm is far more sophisticated than what I’ve described in this crude example, but on the other hand it does seem to have behaved very much like this in Paraguay and some other parts of the SH. (And at Base Orcadas it works in the opposite sense, applying what is clearly a spurious cooling correction. It tries to be impartial, but a flaw in the logic prevents it from doing so.)

          Why doesn’t the algorithm add warming in the NH? I suspect because the vast majority of the records there already show warming. As a result the convergence process provides relatively undistorted averages.

        • manicbeancounter says:

          Further to your comment of April 16, 2015 at 11:22 am
          I believe we need to understand what is meant by “temperature homogenisation.” The data is being adjusted to create regional and global temperature anomalies. Correcting for measurement biases is the first stage of this process. The second stage is to make the data more homogeneous – to effectively reduce the data noise (or eliminate micro climates). This second stage is not understood.
          – The GHCN V2 data might not be fit for purpose. But it has been supplanted by V3 adjusted data. The adjustments are probably increasing over time with respect to older data as they (usually) assume the current data is correct, and adjust backwards. If you have a UHI effect that is increasing, every time a recalculation is done data in the distant past will be cooled more than before. With UHI the opposite should be done.
          – BEST temperature data is unlikely to be imported. from the N Hemisphere. As Roger has described. BEST derive an expected average for a region. The derivation of what is expected might be where errors occur, along with a the highly weighted convergence procedure. I believe the BEST homogenisation methodology is far more rigorous than that of GHCN. Biases that occur will be more consistent, so more worthy of study than the GCHN.

      • Euan Mearns says:

        Is this how you did it Roger? “After break point” minus “raw monthly”? In this case they’ve turned 0.5˚C cooling into 0.87˚C warming? I’ve just had a quick look and the V2 annual Punta Arenas records show about 0.5˚C cooling.

        If so, its quite straight forward to do this comparison for all our stations and quantify the impact of BEST homogenisation.

  4. Euan Mearns says:

    One comparison I forgot to do was with the satellite data. This is UAH, S hemisphere, lower troposphere, land only. I think its pretty good.

  5. Euan Mearns says:

    RA is perhaps even better:

      • Javier says:

        Simply amazing that starting from station temperature records you get the exactly same results as satellite data, while the official databases show so much discrepancy. I am very much impressed. This is the best evidence I have seen so far in favor of satellite data and against official databases homogenization and adjustment.

        Congratulations to both and I hope this info makes the rounds and gets ample diffusion.

        • Euan Mearns says:

          Javier, this is exactly my sentiment. However, I have so far sent this to Andrew Montford, Benny Peisser, Jo Nova, Judy Curry*, Roy Spencer and John Christie and have not yet received a response from any. I set a link on WUWT yesterday that had 4 hits. Maybe its because its Easter?

          * Judy who has run my posts in the past is on Congressional Hearing duty this week.

          I also sent this to Richard Muller and ccd a couple of peers including Matt Ridley. There is a lot of information to take on board.

          Anyway, I sense this may be really important and will persevere . Commenters can of course help by setting links.


          • A C Osborn says:

            Euan, your website has been un-available for the last couple of days, so that may be one reason you didn’t get many hits.
            I doubt you will get much response from the Scientific side, they tried setting their attack dogs on you and that didn’t put you off, so I suspect they will just ignore you.

          • Euan Mearns says:

            AC – its been up full time for me. But interesting to hear that it might have been down for others – not sure how that would work.

            I had a pretty swift response from BEST. Defensive, but at least a dialogue that may lead somewhere.

  6. ducdorleans says:

    very convincing ! …

    independent analysis of data by different persons, and basically the same outcome … (as in UAH and RSS …)

    as to “selling” all this, I can understand the frustration, but I guess this will thus remain … you’re not marketeers, unfunded, don’t have the mass-media’s ear … here in Belgium, even new toilet paper in the offices of Greenpeace, FOE, WWF and other similar charlatans is national news …

    (I humbly propose to also send it to Booker, Delingpole and the offices of the US senator who recently wanted to have an enquiry into these “temperature data institutions” – but unfortunately forgot and cannot find his name atm …)

  7. manicbeancounter says:

    Both you and Roger have done an amazing amount of work here. What I like is that you question your own results, and compare and contrast different methods of analysis, along with different styles of presentation. Roger is extremely good at presenting his results, either by complete posts, or comments that are sometimes worthy of a post on their own.

    Some short comments
    1. I can confirm your comment about the normalisation period not making much difference. When I looked in detail at your Southern African data, I was convinced that it did, so spent many hours trying to demonstrate my belief. It may make a difference with a limited number of diverse data sets.
    2. Your comment temperature data sets being quite close really applies to the variation from one year to another. It is on the more important issue of temperature trends over decades that I believe the divergence emerges. Might it indicate that global temperature anomalies are least reliable on the area which they created for – to measure the longer term climatic changes in temperature?
    3. This brings me onto why I am a “fan” of the 5 year centred averages. The variation from year to year can mask real changes in trend. That is the data noise can visually mask the trend changes. Figure 6 shows the (maybe apparent) trend change from flat trend to warming following the 1976 cooling. It the global data sets the opposite appeared to happen after 1944 and 1998. Upwards spikes seemed to stop warming.
    4. It interesting your note about a lack of evidence for the 1998 El Nino spike, nor for the pause post 1998. This is especially in light of the fact that El Ninos originate in the Pacific in a vast area to the West of Chile!
    5. I would take issue with your comment about station closures in 1990-1991. Rather than physically close, the stations were just removed from the GHCN database. This appears to have been to simplify the homogenisation process. A true regional or global temperature reconstruction needs to take account of the spatial gaps at any given point in time to enable a contour map of temperature trends to be built up. The GCHN method appears to have been a bit of an art form. It meant that some surface temperature stations were completely changed from being a true measurement of a particular locality to being (supposedly) a grid reference point for an irregular area that overlapped with nearby stations.

  8. Pingback: Averaging Temperature Averages | Energy Matters

Comments are closed.