Monday, 28 July 2014

Is the US historical network temperature trend too strong?

Climate dissenters often claim that the observed temperature trend is not only due to global warming, but for a large part due to local effects: due to increases in urbanization around the stations or somehow because of bad micro-siting.

A few days ago I had a twitter discussion with Ronan Connolly. He and his father claim that 0.2°C per century of the temperature increase in the USA is due to urbanization and 0.1°C per century is due to micro-siting. That is quite a lot. Together it would be almost half of the temperature trend seen in the main global datasets.

One of the great things of America is that they have a climate reference network (USRCN). The observations are normally made by the meteorologists and contain non-climatic effects that are not relevant for the meteorologists, but they are to climatologists. Thus to track accurately what is happening to the climate, NOAA has set up a climate reference network that follows high climatological standards. The main thing for this post is that these stations are located in pristine locations, without any problems with urbanization and micro-siting.

We only have data from this network starting in 2005. That is only a decade of data, but if the problems with the normal data are as large as Connolly claims, I thought we might be able to see some differences between the reference network and the normal US historical Climate Network (USHCN). In the USHCN non-climatic effects have been removed as well as possible with the pairwise homogenization algorithm (PHA) of NOAA.

The figure of NOAA below (in Fahrenheit) shows that USHCN (normal) and USRCN (reference) track each other quite closely. If you look at the details, you can see that actually the USRCH is a little below USHCN in the beginning and a little above at the end. In other words, the temperatures of the reference network are warming faster than those of the normal network. The opposite of what the climate dissenters would expect.

Let's have a more detailed look at the difference between the two networks in the following graph. It shows that the warming in the reference network was 0.09°C stronger per decade. For comparison with the trend due to global warming, you could also say that it is 0.9°C stronger per century. That is just as much as the observed global warming trend.

That the trend in the normal data is an underestimate of the true warming is no surprise for me. The trend of the raw American data has a strong cooling bias. Removing of non-climatic effects (homogenization) increases the temperature trend since 1880 by 0.4°C. We also know that homogenization can make trend estimates more reliable, but cannot fully remove the bias. Thus it was likely that there was a remaining cooling bias.

The cooling bias could be due to a number of effects. An important cooling bias in the USA is the transition of conventional observations with a cotton region shelter to automatic weather stations (maximum-minimum temperature systems). This transition is almost completed and was more intense in the previous century. Other biases could be the relocation of city stations to airports. This mainly took place before and during the second world war. The increased in interest in climate change may have increased interest in urbanization and micro-siting, which may thus have improved over time due to relocations. (Does anyone know any articles on that? I only know one for Austria.) There is also a marked increase in irrigation of gardens and cropland the last century.

That the effect is this strong is something we should probably not take seriously (yet). We only have nine values and thus a large uncertainty. In addition, homogenization is less powerful near the edges of the data, you want to detect changes in the mean and should thus be able to compute a mean with sufficient accuracy. As a consequence, NOAA does not adjust the last 18 month of the data, while half of the trend is due to the last two values. Still an artificial warming of USHCN, as the climate dissenters claim, seems highly unlikely.

This cooling bias is an interesting finding. Even if we should not take the magnitude too seriously, it shows that we should study cooling biases in the climate observations with much more urgency. The past focus on detecting climate change has led to a focus on warming biases, especially urbanization. Now that that problem is cleared, we need to know the best estimate of the climatic changes and not just the minimum estimate.

Maybe even more importantly, it shows that we need climate reference networks in every country. Especially to study climatic changes in extreme weather in daily station data, data that is much harder to homogenize than the annual means. We are performing a unique experiment with our global climate system. Future scientists will never forgive us if we do not measure what is happening as accurately as we can.

[UPDATE. In case anyone wants to analyse the "dataset", here it is:
diff = [0.017 0.050 0.022 0.017 0.022 0.017 -0.006 -0.039 -0.033]; % Difference USHCN-USRCN in °C
year = 2005:2013;

See also

Howard J. Diamond, Thomas R. Karl, Michael A. Palecki, C. Bruce Baker, Jesse E. Bell, Ronald D. Leeper, David R. Easterling, Jay H. Lawrimore, Tilden P. Meyers, Michael R. Helfert, Grant Goodge, Peter W. Thorne, U.S. Climate Reference Network after One Decade of Operations: Status and Assessment, Bulletin of the American Meteorological Society, 2013, 94, 4, 485, doi: 10.1175/BAMS-D-12-00170.1.


Steve Bloom said...

IIRC the CRN stations have not merely good siting, but triple-redundant ambient thermometers and a downward-facing IR thermometer. Do you know anything about adjustments involving the latter?

Victor Venema said...

Hi Steve, yes another feature of these stations is that they are triple redundant for most of the measurements. Furthermore, the data are send to NCDC in real time. Thus if one of the instruments differs from the other two, a technician is send to investigate and fix the reason.

The USRCN data is not adjusted. Infra red radiation from the hot soil is an important warming bias source is older measurements. For example, Manola Brunet et al. performed a comparison of a Stevenson screen with a French screen. The French screen is open to the bottom and had a considerable warm bias of more than 1 degree Celsius as a consequence. See slide 11 of this talk.

Peter Thorne said...


the triple redundant sensors are all NIST calibrated so no adjustment should be required. As Victor said this really should be something we have globally (not nec. at the density of the US network). There are a number of us working hard to make this a reality and the direction of travel is at least encouraging. Its far from a case of snap one's fingers and it magically appears, sadly.

protonsforbreakfast said...


Thank you for this post. I am currently jet lagged and following advice in your previous post I am working while I can.

I too am an admirer of the USCRN and – as with the ARGO floats – it shows that although new initiatives take a lot of resource to start, they begin to pay back on timescales as short as a decade. And on a time scale of a century they become critically valuable. I am sure people in 100 years will look and back and ask why it took us all so long. And perhaps they will also ask why there is no humidity measurement in this reference set.

Onto the topic of your post, I overcame my horror at your use of the F-word and found the agreement between the two networks was surprisingly close – and thus reassuring.

However your post has my jet-lagged head aching and indicates that perhaps I don’t understand what you have done. My concern relates to the fundamental measurement uncertainty of the USCRN. Despite poking around for an hour I could not find a statement of uncertainty of measurement. But we can make a guess.

The raw resolution of the datalogger is 0.01 °C (15 bit resolution) of a 1000 ohm PRT.

There is no discussion of measurement current or self-heating and given the fineness of the wires in a PT1000 sensor and our own experience with these sensors at NPL, I find the claim of 0.01 °C/year drift surprising on the optimistic side. It would be interesting to see how the calibration shifts from year to year. Keeping a reference resistor stable at this level is also no simple task.

Overall, raw measurement uncertainty is probably u(k=1) ≈ 0.03 °C. Measuring it 1000 times never improves this uncertainty of measurement.

But in a raw comparison between two climate networks – one established and the other new – it seems to me that it is impossible to meaningfully compare numbers at a level better than the basic single sensor uncertainty of measurement. If the second network has similar or larger uncertainty of measurement the effect (if uncorrelated) is multiplied by square root 2.

So I think on your second graph you should have one-sigma error bars of around 0.05 °C and that in the face of such uncertainty, and trend is unlikely to be significant.

Please let me know if you think I have missed the point.

Every best wish

Michael de Podesta

Victor Venema said...

Michael, I hope I was so careful not to claim that the negative trend in the difference is statistically significant. You probably have to leave out the last two points and take auto-correlations into account. Then the negative trend is not significantly different from zero. For such a short time series, it is very hard to distinguish auto-correlation from a trend.

I only wanted to show that there are no signs of a large warming of 0.3°C per century in USHCN.

Do I understand it right that your claim is that the drift of PT1000 sensors is systematic, not random? There are 100 USRCN stations, thus if the drift is random, the influence of such a drift would be 10 times smaller.

I fail to see how the raw resolution is important. At least here, measuring it 1000 times (and in this case a lot more) certainly improves the uncertainty of the estimate.

Ronan Connolly said...

Hi Victor,

First off, apologies for not commenting sooner in your post directed at me. As I mentioned on Twitter, I was moving house on the day you posted this & the last two weeks have been pretty hectic, so I’ve only really got back to checking Twitter & the blogs since Sunday!

Anyway, a few comments:

1. “Climate dissenters”? What exactly is a “climate dissenter”, and how are me & Michael ones? Sometimes, the weather will “disagree with me” (Ireland has a lot of wind & rain unfortunately!), but I don’t think I’ve ever disagreed with (or “dissented” from) the weather (or climate)! ;)

In all seriousness, I know it’s very difficult coming up with a simple label to describe people you disagree with in the climate debate that don’t cause offence. But, there’s a reason for this!!!
Simplistic labelling divides people into “us vs. them”, and if you then imply that “your” group is the “right” group, you’re automatically disrespecting and dismissing “them”. This is going to happen whichever term you’re going to use to describe “them”, and this just furthers the divide.

In my opinion, the climate debate has already become far too polarised and divisive, and anytime anyone uses an “us vs. them” label it just makes the “climate tribalism” worse. Do you not agree?

Is it really necessary to apply a generic label to all those people in the climate debate who disagree with you?

On our Global Warming Solved blog we try to avoid this labelling. If we’re describing an aspect of somebody’s work that we disagree with (and it’s relevant for the discussion), we might say so (& briefly mention why we disagree). But, we’re more interested in the science than in “picking sides”. Aren’t you?

Obviously, there are aspects of climate science that you and I disagree with each other on, e.g., the extent of urbanization bias in global temperature trend estimates. But, isn’t that always the way in science?

If you still want to use a label to more accurately describe me & Michael in terms of our urbanization bias research findings, instead of using the silly & nonsensical term “climate dissenters”, I guess you could say something like “some scientists who have come to a different conclusion on the magnitude of the urbanization bias problem through their scientific research than I have”. ;)

But, why bother? Isn’t it simpler just to say “In their series of papers studying urbanization bias & siting biases, Drs. Ronan Connolly & Michael Connolly argue that the magnitude of these biases is greater than is currently assumed.”, or something similar?

2. Although the USCRN does indeed have more than a decade of data, it is important to realise that most of the stations in the network still have much less than that:

Year, USCRN stations with full year’s data, % of current stations
2001, 2, 1%
2002, 8, 4%
2003, 25, 11%
2004, 45, 21%
2005, 72, 33%
2006, 82, 38%
2007, 97, 44%
2008, 121, 56%
2009, 137, 63%
2010, 155, 71%
2011, 201, 92%
2012, 218, 100%
2013, 218, 100%

Only 1/3 of the stations were actually operating for the full 2005-2014 period you’ve shown, and 29% of the stations were only set up in the last five years.

[To be continued...]

Ronan Connolly said...

[part 2 ...]

3. The NOAA plot you show claims to describe the “national departures from normal” for the USCRN, i.e., the mean annual anomalies. But, this is a bit misleading because NOAA usually define climate “normals” as the average over a 30 year period. If some of the USCRN stations only have 2-3 years this isn’t enough to calculate the 30 year average for that station!

Instead, according to the NOAA NCDC website where you got your data from (NOAA NCDC), to approximate the normals for each of the stations, they calculated the 30 year mean temperature from the COOP neighbours (following Sun & Peterson, 2005’s approach - here).
I can appreciate the logic of this approach – they’re trying to make the most of the data they currently have. But, when you combine this with the changing number of stations in the USCRN, do you see how this isn’t good enough to justify the conclusions you made in this post?

For instance, each year, whenever a new station is introduced into the network this will alter the national “normal” and average “anomaly”.

4. Initially I assumed that NOAA NCDC were taking care to ensure that they were locating the USCRN stations far away from urban areas, as well as making sure their local siting was of a good quality. Indeed, in the Recommendations section of our “Has poor station quality biased U.S. temperature trend estimates?” paper (here), we actually suggested that networks like the USCRN could ”...if properly managed, have the potential to offer future researchers climate records which would be unaffected by many of the main non-climatic biases plaguing current studies, e.g., urbanization bias, siting bias, time-of-observation bias and possibly instrumental bias”.

However, in light of your post, I’ve had a closer look at the station locations using a Google Earth file of all USCRN 218 stations I created as Supplementary Information for our UHI Paper 1. I’ve uploaded a link, in case you (or your readers) use Google Earth & are interested: here

I’m now not so sure they’re doing as good a job at avoiding UHI as I had assumed. :(

They do seem to have located most of the stations outside the urban boundaries. But, in some cases, they’re only a couple of kilometres away. It is quite possible that these stations will become surrounded by urban sprawl over the next few decades – reintroducing the UHI problem yet again. Even already, looking using Google Earth, stations such as Santa Barbara, CA might be affected.

5. Having said all of that, I want to stress that, like you and Peter Thorne, I also think similar networks should be set up in other countries a.s.a.p. In fact, this was our final recommendation in the Surface Stations paper I mentioned above (see p31, lines 2324-2341 - here).

Even if it doesn’t entirely get rid of the UHI problem, if they are properly maintained & monitored, this should reduce many of the other non-climatic bias problems, e.g., siting biases, TOB biases, station move biases & instrumental biases.

Not only that, but having access to 5 minute-interval measurements for every day of the year is far more useful than a simple monthly mean of “T_max+T_min/2” values! If we could have a similar network for other countries, this would be of tremendous value for studying “short-term climate variability”.

Victor Venema said...

Hi Ronan, a communication expert advised to use the word climate dissenter as the most inoffensive possible option, because as he said, these people definitely value their dissent. Sometimes I have the feeling that nothing short of "Bearers of the Holy and Immutable Truth that climate science is fundamentally and completely WRONG" would do. And, I am sorry, but that is not an option for me. This political correctness is really tiring.

Anyway, if you reread the text, you will find that I did not refer to you as a dissenter. Had this post been just an answer to you, I would have send you a mail and would not have published this at a time I knew you were busy. However, next to you there are many more people that hold the view that a large part if not all of the warming is non-climatic.

In referring to a group, you need some sort of label, I cannot list all names. The US Tea party is too big. Groups do exist, even if their boundaries are not always well defined and individuals will often feel wrongly classified.

I also have a generic label for people I disagree with: humans. I am somewhat of a contrarian by nature and it is my job to find points where I can legitimately disagree with other scientists. Given that it is my job to criticize climate science, it is important for me to distinguish my legitimate critique (for example on the quality of daily data for trend analysis of extreme weather) from the ill-informed and thinly veiled utter nonsense that is typically presented on WUWT and Co.

If you want to be seen as part of science, I would advice you to do the same. (And I would submit your manuscripts to independent scientific journals, rather than fooling your readers into thinking that these manuscripts are part (or will soon be part) of the scientific literature.)

Back to the science of this post and your comments. I did not come to any clear conclusions. For that I would need uncertainty estimates. If I had those, I would not write a blog post in summer, but write a Nature article. One of those uncertainties is the gradual build up of the USRCN is a potential problem. Also small differences in the kind of climate represented by USHCN and USRCH may be a problem. Both these problem could be studied by comparing the USRCH stations individually to their nearest neighbor in USHCN.

These points, however, like the others already mentioned above are sources of uncertainty, not of a bias in the trends. In case you were right, one would have expected that USHCN is warming faster than USRCH. Thus it does suggest that it is more likely than not that you are wrong. However, without uncertainty estimates, I cannot come to the conclusion that you are wrong. It is possible that the there is still a sufficiently large chance that you are right. And we should also not forget the other evidence, this is just one line of evidence.

Thank you for the Google Earth link, that is very helpful. The link seems to be broken, the right one is here. I had also thought that they installed the USRCH stations at locations that would not become urban in the coming century. If you have a list of stations that you think do not fit to that rule, I could imagine that NOAA would be interested. If you have such a list please send it to me or Matthew Menne. One station in Ashville also looks relatively close, but I am no geographer, no idea how much further one can expect Ashville to expand, where it is possible to build.

And thanks for your support for a global climate reference network. Had your manuscripts been published in the scientific literature, ;-) I could have cited that in my upcoming papers.

EliRabett said...

As Eli recalls, the USCRN was a pairwise design with nearby USHCN stations and in that case the interesting point was that although absolute temperatures varies, the variation over very short periods of time were almost exactly the same

Victor Venema said...

I thought the same, but this bunny learned better this week. There are only 13 USRCH stations near USHCN stations (within 500m). Jared Rennie has a manuscript about a comparison of these station submitted. I hope it will also study the questions about the above post, but I do not know the details yet.

Anyone interested in this list of 13 stations can mail Jared or me.

EliRabett said...

Don't think that it had to be totally pairwise to achieve the needed result tho.

Zeke said...

We also had a paper in JGR on this subject (urban-correlated trend biases in USHCN) that might be of interest Rohan. We found a that about 20% of the century-scale trend in the min raw data could be explained by urban-correlated biases, but that these were effectively removed by homogenization, even when the PHA was run such that only rural stations were used to detect and correct breakpoints.