At the American Meteorological Society conference in Asheville last month I presented the initial results of an ongoing project by Matt Menne, David Jones, Ron Broberg, Troy Masters, Claude Williams, and myself. The AMS just posted an audio version of the talk (with slides), for those interested.
You can also find a it at the AMS site.
Other Interesting AMS Presentations:
Unfortunately, you can’t hear the questions at the end, but John Christy asked a question about the effects of homogenization distorting or removing the actual warming occurring in cities, and how that might be problematic for calculation of local temperature trends.
The second question (I’m not sure the name of the person who asked it) involved taking station moves into account, particularly in the earlier part of the century.
Taking station moves into account is theoretically an issue to that I imagine the Menne et al presentation addressed, or is based on the USHCN V2 algorithm which is supposed to try and account for such things.
Anyway, it’s good to see you’ve presented your work to the wider scientific community. It will get more attention that way.
Nice presentation, Zeke. Wish they had video.
Carrick
Andrew_FL,
As I mentioned in my response to the question (near the end), the USHCNv2 homogeneity process uses all available data to help detect inhomogenieties. Specifically, when metadata suggests a station move, tobs change, or instrument change the significance threshold used to flag inhomogenieties is lowered relative to that used to flag undocumented issues. Unfortunately, however, the metadata itself isn’t always correct (Menne touches on this a bit in his talk), so its not always that simple.
On Station Moves.
I think its an interesting excercise to imagine how station moves can make a difference.
1. You take an network that is All rural by a proxy at the end of the series.
2. That rural series shows the same warming trend as an urban network.
What are the concerns about the rural series?
that moves were made in early periods from warmer places to cooler places, thus creating an artifical warming trend in the rural series. That could happen for these reasons
1. deurbanization
2. moves from the urban to the rural
3. Moves from lower stations to higher stations
4. Changes in latitude. (hmm how much?)
I think we are at a stage where concerns like this need to be made concrete. That is, show some examples.
Steven-a station move can have the effect cooling or warming the station not just for concrete, obvious reasons, but simply because of seemingly random variations in microclimate. Since it can go in either direction, one cannot say ahead of time how it might influence the detection of urbanization. Movement of stations has to be dealt with seperately, ideally I would think deal with such effects first, then do rural and urban comparisons, but as long as the effects are random and normally distributed around zero and have no temporal instability overall in terms of frequency, they only mean uncertainty is greater for rural-urban comparisons that haven’t been pre-corrected for spurious shifts due to moves.
“Steven-a station move can have the effect cooling or warming the station not just for concrete, obvious reasons, but simply because of seemingly random variations in microclimate.”
What I am getting at is this. If Zeke shows a history of rural stations, that start at 0C anomaly and end at 1C anomaly, and you raise the issue of station moves, the obvious question in my mind is this. The sum of all the moves can have 1 of three effects:
A. No effect since they balance out.
B. A cooling effect
C. A warming effect.
So which are you concerned about and how would that effect happen. be precise.
Obviously you try to account for all the documented moves you have. So, is your concer for documented or undocumented. So if you comment is just “hey look for station moves” then well, of course. If your concern is about undocumented moves, then The following logic applies:
If your concern is that rural stations NOW, might have been cooler earlier ( a warm bias move) Then that’s one thing. If your concern is that rural now now might have been warmer earlier ( cold bias moves— see my examples) Then thats another thing.
So which concerns you more and why?
Interesting enough, a lot of currently rural stations (especially airport ones) moved out of urban areas post-WW2. That said, lots of the city airports that started out as relatively rural have become more urban over time, but there is still a difference between an airport-sited station and one of the roof of a building (a common case pre-1940s).
After listening to the linked video above on the ubiquitous UHI warming, I would hope that the investigations of temperature and temperature trend uncertainties could now be more focused on the micro climate citing of temperature measuring stations. I heard some statements early in the talk that sounded encouraging in those regards.
I realize that looking in more detail at station micro climate changes over time and its potential effects on the trend uncertainties is a far more daunting task than looking at population changes and satellite night lights and UHI effects, but I sometimes get the idea that we look for these uncertainties like we do for the lost keys under the street light because it makes the searching easier. That is not to say that these studies such as noted in the video are not of relevance.
On further thought I would guess that having the break point algorithm of USHCN might give some confidence for catching micro climate changes at a reporting station over historical times. That confidence would have to come with some experiments/models that would simulate various types of change and then applying the algorithm to those changes to determine whether the method could capture the change magnitudes – given that some changes no doubt occurred slowly over time.
Has any work such I mention here been initiated to date?
Kenneth,
The Menne et al talk on testing homogenization linked at the bottom of the post discusses testing the break point algorithm on various synthetic data sets. Its definitely worth a watch.
steven mosher (Comment #80323)-If you are asking me personally, I am not more concerned with cooling moves or warming moves. I only want this to be carefully taken into account, independent of the urbanization issue. Not just the rural sites could be moved by the way but also the urban ones, moved to different parts of a city, for instance, where the intensity of the heat island may be different, among many other things.
So a priori I tend to think these things mostly cancel out. The concern would be, what if they don’t? People who are not me, incidentally, probably have different concerns, probably are inclined to believe the station moves must not cancel out and must create a warm bias. I know it is hard for you to believe, but those people aren’t me, and I don’t think that. I don’t know what to expect. I just want to know whether this matters or not. Okay, here’s a hypothesis for you Steven, from me: in the US at least I am prepared to say that the net effect of station moves on the assessment of UHIE will not be statistically significantly different from zero in either direction.
Nice presentation, Zeke.
Excellent job Zeke!
By the way, folks, the AMS meetings are always interesting in general and there are loads of presentations on other issues. I recommend checking more of them out:
http://ams.confex.com/ams/19Applied/webprogram/19APPLIED.html
Thanks for lead on the Menne talk, Zeke. I’ll have a look.
Anyone read the paper by Chen et al in the latest J Geophysics Res. which says up to 44% of the temp increase in eastern China in the last 30 years is due to urbanisation (UHI)?
Zeke, I listened to the Menno video, but before I comment I want like to compliment you and Menne on your presentations for the ease of following what you were presenting.
I need to go over the details of the Menne presentation, but I judge it to be a good start to testing the breakpoint algorithm. The results, good and not so good, were presented and without spin. It is also commendable that he covered Type I and II errors.
What I was looking for was a simulation of citing changes as has been detected by the Watts team observation. I heard Menne refer to a question about instrument issues and deferred to a later presentation and he seemed to be saying that instruments caused bigger changes than those from citing. I’ll have to listen to that presentation next.
Re: Kenneth Fritsch (Aug 12 14:46),
Sorry, can’t resist.
I presume you mean siting (location) rather than citing (literature references, e.g.).
At least you’re consistent. Spell check doesn’t work when the misspelling is a word too.
Dewitt Payne, thanks for the insightful findings of my misuses of citing, not once but twice. Even more revealing after my crusading on another thread about quality control. I was going to say that I distinctly heard Menne pronounce the c and not s in siting, but decided it best to admit to a rather dumb mistake.
Actually I do not believe that siting is a word, and, in fact, I should have been referring the effects of micro sites.
It’s a word.
It wasn’t a word by my spell checker and that is when I changed siting to citing without thinking. The Webster dictionary (and Carrick) cites siting as a verb and that is good enough for me.
I listened to the station siting video which was a rehash of a summary that you had posted, Zeke. I need to answer my own questions on what the CRN rated station comparisons look like as one goes back in time and uses CRN123 versus CRN45 stations. All the CRN data used in the Fall paper are available and I just need the time to process it.
Upon sighting the site I decided to cite some who had sighted the site earlier. 🙂
I’ve noticed that many spell checks are not up on all these fancy scientific terms. “Forcing” is another one. Until about two seconds ago I was sure that it was Firefox that kept complaining about that word. Thinking about it, I think it’s WordPress and MS Word that give me all the trouble with that one. Hm… irradiance is not recognized by Firefox as a legitimate word… insolation isn’t either. Yup. Take care in climate science, you’ll have to learn to spell these terms watching other people. Many spell checks have never seen these terms before!
Thanks Andrew.
So back to my point. If a person is concerned about an all rural series being contaminated by station moves that bias it cool in the early years… then, what types of things does one need to look for?
1. deurbanization
2. moves from the urban to the rural
3. Moves from lower stations to higher stations
4. Changes in latitude. (hmm how much?)
so. I think there are ways to address those concerns. But I dont want to engage people and answer those 4 questions if they are going to make up 6 more behind that.
From Menne’s presentation, I was surprised that a climate model would have the necessary resolution to simulate station location temperatures.
My old unabridged Webster dictionary shows cited and citing under cite but not sited and siting under site. Time to throw that old dog-eared heap away.
My 1971 compact edition of the Oxford English Dictionary lists both sited and siting.
Re: Kenneth Fritsch (Aug 13 07:58),
In the opening scene of the Nero Wolfe story “Gambit”, Nero Wolfe was tearing pages out of the latest edition of “Webster’s International Dictionary, Unabridged” and tossing them into the fire. What was the reason for Wolfe’s apparent eccentric behavior? He thought it was guilty of deliberately murdering the English language.
Andrew_FL:
That is a bit …strange. “Forcing” is the transitive form of “force” as to “shove something open.”
“Stop forcing the door, you’re going to break it!”
Using verbs as nouns (when they end with -ing) is also standard English usage.
“Rick Perry has announced he will be in the running for President”.
steven mosher:
I think the changes in latitudinal distribution probably matters the most, but only then if you have missing cells in places that have undue weight (the far north), and can’t simply replace the missing cells with the average of surrounding cells.
It leads to an increased uncertainty, and a north/south trend in geographical coverage can in principle lead to an anomalous warming or cooling. But only then with the caveat of missing cells.
Carrick (Comment #80357)-Checking again, I think I was mistaken. It is the word “radiative” that gives MS Word, at least, trouble.
Carrick (Comment #80358)-I think this is why pairwise comparisons are normally used.
Re: DeWitt Payne (Comment #80356)
(Nero Wolfe vs. Webster’s)
I haven’t read that story for a while. Was that because it used “contact” as a verb, or because it claimed that “infer” and “imply” could be used interchangeably?
Andrew_FL, pairwise comparisons only help if there is more than one sensor in a cell. If you are spatially undersampling a region there isn’t any real cure for it—you’ve simply lost information, no easy way to recover it.
This undersampling is more important in the northern most latitudes, since those have a larger contribution (intrinsic weight) in global mean temperature trend. Unfortunately that’s where some major holes in the grid are found (I’m thinking of Siberia, not Canada for the record).
Re: Julio (Aug 13 11:52),
Probably both.
I didn’t actually look up the story. I just googled ‘Nero Wolfe burning dictionary’ because I remembered the scene but not the details.
Can you imagine his reaction to using impacted as a transitive verb rather than hit. It seems to have become standard usage, but it’s really annoying.
Carrick (Comment #80364)-The under-sampling issue (heh, Firefox says that’s not a word either…stupid spell check) is why I am not particularly worried about the data from the US, which is very well sampled. In terms of Urbanization I wouldn’t worry that much about Siberia. It’s pretty undeveloped and sparsely populated. Also there is probably a good physical reason for the strong winter warming in the Eurasian Anticyclone:
http://www.ottokinne.de/articles/cr/14/c014p001.pdf
For under-sampling I worry more about Africa, even though the trend contribution is not that great. Although it may not matter much for global averages, I prefer to know for sure than to assume it won’t matter.
Re: Andrew_FL (Aug 13 14:37),
You know you can add words to the Firefox dictionary if you’re certain of the spelling. Right click and select ‘add to dictionary’. I usually open a new tab and google the word if I’m not entirely certain of the spelling rather than getting out the compact OED and magnifying glass.
DeWitt Payne (Comment #80371)-Thanks, I’ll probably do that, I just usually can’t be bothered.
Re: DeWitt Payne (Comment #80367)
You’re right, he would not have approved of the transitive “impacted” either, or the transitive “grow” as in “grow the economy” (“grow tomatoes” is all right, of course, but it has a different meaning).
Then again, that is how languages evolve… as long as they stay alive it is probably unavoidable.
Carrick.
Latitude changes.
A long time ago ( in blog years) NCDC used to have a chart showing the changes due to station moves, from a program called SHAP.
( hmm I prolly can find it) SHAP estimated offsets for changes in alt ( prolly simple lapse rate) and latitude/long.
The issue is undocumented moves and or wrongly documented moves. nevertheless, the sample of documented moves could give you some insight into modelling undocumented moves.
We already have one boundary. We assume no undocumented moves . what do the other boundaries look like?
It seems to me that a good test of a homogenization algorithm, statistically, would involve first generating fields of noise time series (probably red, but how red?) and the different series should correlate spatially (how strongly, at what rate of distance drop off for correlation?) and then introduce at random jumps and trends to some of the series, (with what frequency and magnitude relative to the magnitude of the noise?) and see if the algorithm developed correctly identifies the discontinuities so introduced. Of course, the problems with doing this are developing reasonable models for both the kind and level of discontinuities, and the properties of the noise model. These will actually probably determine what the optimum parameters are for a inhomogeneity correcting algorithm. One can of course pick some random set of parameters, but whether these bear any resemblance to the real world is a serious issue. Perhaps climate model simulations could produce fairly realistic bases instead of a noise model, realistic “stations” would probably be model gridcells plus some level of white noise. Realistic levels of discontinuity is a more difficult problem. Of course, one can also use such an approach to see how sensitive one’s algorithm is to density of stations. So conceptually I can see a lot of difficulties constructing a good test of any homogenization algorithm. How such difficulties might be overcome eludes me, however. Any ideas?
Stephen, here’s my computation for CRUTEMP for change in mean latitude of station (using a gridded algorithm),
and it’s effect on mean temperature trend (relative to an Earth that is all land, so the roughly 7% remaining bias is intrinsic to the fact there is more land in the northern hemisphere than the southern.)
I take it they didn’t give Zeke much time to make his presentation, he talked so fast that he came off like a salesman trying to pull the wool over someone’s eyes. It’s good to see the guardians of the temperature record finally address the UHI issue… particularly since Parker and Peterson couldn’t find it in their bogus studies in the last IPCC report.
What concerns me in this work is the lack of addressing the automated algorythem issue where the discontinuities are missed by an average of .4 … a pretty phenomenal number when the signal claimed to have been detected was .6.
Lots of presenters at scientific meetings talk fast to fit what they want to say into the time limit.
… so much for aesthetics… anything on the real issue?
“he talked so fast that he came off like a salesman trying to pull the wool over someone’s eyes”
http://www.youtube.com/watch?v=GZxsR8DaPnY
Andrew
“Using verbs as nouns (when they end with -ing) is also standard English usage.”
Quite true and makes me wonder why my old dictionary does not show sited and siting. It is missing its cover and first several pages so in attempt to date it I went the section of US presidents. The last one listed was Jimmy Carter. Nero Wolfe would never own a dictionary as ratty as mine and if he were discarding it into the fireplace it would be because he would ashamed to be seen with it.
Andrew_FL (Comment #80377)
I would think that if a breakpoint analysis could work (using an optimum algorithm) with good confidence that it would provide a way to analyze changes from historical times concerning such as issues as Watts CRN ratings – that do not provide when a change occurred.
Breakpoint analysis that I have done would appear to be not entirely objective in that minimum segments are left to the discretion of the user and there is always the issue, I think, of breakpoints within a breakpoint segment.
It does remind me that, when I do my analysis of the Watts CRN ratings, breakpoints will be included.
“What concerns me in this work is the lack of addressing the automated algorythem issue where the discontinuities are missed by an average of .4 … a pretty phenomenal number when the signal claimed to have been detected was .6.”
Thats the old approach MikeC. A couple points.
a .4C step discontinuity in a long series is not that big of a deal when it comes to the trend calculation, actually it depends on where the discontinuity appears and in how many series. In any case you’ll have to wait to see the final work and then you’re old talking point ( which relative of yours explained this to you again? ) will kinda be worse than the gossip it is now.
Kenneth Fritsch (Comment #80418)-I think the only solution to subjectivity in these or any other analyses is to increase the number of analysts. By the central limit theorem a large enough number of people converge on the truth 😉
That last part was a joke, just to be clear.
Anyway, it’s good that NOAA has a breakpoint algorithm for USHCN, I think. Radiosonde products that have been constructed by various groups and pretty much all of them rely on some kind of breakpoint detection to produce homogenized time series. I think it is quite worthwhile to use these approaches with other types of climate data, and the way I see it this has not been adequately examined as an issue with regard to global surface temperature data.
“I think it is quite worthwhile to use these approaches with other types of climate data, and the way I see it this has not been adequately examined as an issue with regard to global surface temperature data.”
Agreed. I would like to see more analyses of the breakpoints in the historical temperature records of the globe and regions of the globe. It would be even better if the analysis were accompanied with an explanation of why we should see linear segments. It is rather obvious, at least to me, that we do not have a linear trend over the entire time period – as Lucia has often expounded upon here at the the BB. I would guess that many observers would say there are good reasons to suspect that the trend is non-linear, although I have not seen anyone seriously propose an equation. Next best might well be segmented linear trends.
I would admit though when I did some breakpoint analyses in R and decided to do a breakpoint analysis on the derived linear segment and got more breakpoints I was a little disappointed since I did not have a good explanation for the originally derived breakpoints. I felt like I was being fractaled.
Kenneth Fritsch (Comment #80435)-I think the idea with break points involves expecting linear trends and jumps in the differences between nearby stations. The reasoning is, I think, that the climatic trend shared between neighboring stations should share all the same influences, the same effects that give rise to climatic shifts, oscillations, and trends, should be approximately the same for stations located very near to one another (because anomalies are supposed to be large scale) so the differences between neighboring stations would be expected to be white noise with amplitude much lower than the variations in the actual temps at those stations. The presence, then, of a trend or a shift in those differences is thought to be indicative of a non-climatic shift in one of the stations’ timeseries.
Andrew_FL:
Pretty much. If you haven’t read the Menne and Williams 2009 paper on pairwise homogenization, the paper is free to access over on the NCDC site: ftp://ftp.ncdc.noaa.gov/pub/data/ushcn/v2/monthly/menne-williams2009.pdf
Yes, I understand that differences are the key to using breakpoints of difference series between adjacent stations and as such it does put the method closer to having a prior rationalization than when using breakpoints for analyzing stand alone historical temperature series.
What I was thinking about was looking back in time for breakpoints associated with the Watts CRN ratings. For example, the Menne and Fall papers confined their analyses of CRN rates stations to the 1980s to present and that time period would miss the effect on trends of micro climate site changes that occurred prior to that time period. Now we have a method (using breakpoints of difference series of adjacent stations) that USHCN claims can efficiently find site changes, yet no one that I am aware of has used that analyses when looking at CRN rated stations in efforts to determine when the siting changes might have affected temperature trends.
There is a conundrum in that the USHCN algorithm using breakpoints appears to see most of the homogenization adjustments coming from instrument changes and not micro climate site changes. Perhaps micro climate site changes occurred too slowly for the breakpoint settings used to find instrument changes and perhaps a more sensitive analysis is required to find micro climate changes.
Well, the CRN ratings are (in their current form) just snapshots, and can’t really give you much historical information. The metadata does have information on station moves, and the pairwise algorithm uses that to help identify breakpoints. Are you suggesting looking at current poorly-rated CRN stations more closely for breakpoints than current well-rated CRN stations? Bear in mind that a number of well-rated CRN stations moved out of city centers to airports in the 40s, so there are moves there as well.
“Are you suggesting looking at current poorly-rated CRN stations more closely for breakpoints than current well-rated CRN stations? Bear in mind that a number of well-rated CRN stations moved out of city centers to airports in the 40s, so there are moves there as well.”
As I think Lucia would reply: Why would you think that?
What I had in mind was looking at all CRN rated stations with a more sensitive algorithm to determine if micro climate changes might be revealed and, of course, the time of the change. Certainly meta data could be used to avoid confusing micro climate changes at a given site with site changes.
I well might find that breakpoint analysis will not provide an unambiguous picture of micro climate changes or that indeed micro climate changes have much less effect on the temperatures and temperature trends than I would expect knowing that the micro climate differences with location in my yard can be very noticeable.
I am currently finishing my analysis of the Mann 08 temperature reconstructions and have discovered that actually digging in and looking at the base data gives a very different picture than that presented in the paper. In the case of site micro climate changes I simply think the work to date is incomplete. Unfortunately, since papers do not normally mention or dwell on avenues of investigation that appeared unproductive to the authors, I may be treading on the same ground again.
Kenneth Fritsch,
I guess my confusion arises from the fact that micro climate changes are not documented per se (we only have a current snapshot). The closest proxy we have, metadata recording station moves, is already used to calibrate the breakpoint analysis such that the threshold for inhomogenity detection is lower for documented changes.
One problem for climate studies is that adjustment procedures are often developed with monthly averages of max and min temperatures in mind. This means, basically, that while we can do studies of mean temperature trends and even seasonal breakdowns and at least say we think we have adequately accounted for the impacts of various non-climatic effects, the study of extreme climate events on daily timescales is much murkier. Does anyone know if anyone has made attempts to develop adjustment algorithms that would potentially work on sub monthly timescales? Or is such a thing impossible?
Andrew_FL,
I know that there are some efforts underway to apply breakpoint detection and homogenization to daily temperature data (both at NCDC and via the Berkeley Earth project). I’m not sure if anyone is planning on tackling hourly temperatures yet, however.
Homogenizing daily data vs. monthly data won’t change long term trends much, but will help with tracking changes in extremes and whatnot. There are probably some interesting issues introduced by CRS to MMTS conversions, particularly for hourly temperatures.
Zeke (Comment #80468)-Thanks, that’s basically what I thought. Personally, I have a lot of interest in the issue of tracking extremes, and I would like to see work dealing with homogeneization of daily scale data whenever it gets published. The sub daily scale, ie hourly might be interesting, too, although personally I doubt that it will be of much more than academic interest in terms of fine detail of the data.
If Mosher thinks .4 misses are not important in a formula which finds a .6 trend then I’d say smoke a little less crack in the bath house…
Zeke (Comment #80463)
“I guess my confusion arises from the fact that micro climate changes are not documented per se (we only have a current snapshot). The closest proxy we have, metadata recording station moves, is already used to calibrate the breakpoint analysis such that the threshold for inhomogenity detection is lower for documented changes.”
I am not sure what you are saying here, Zeke. It is my understanding that USHCN adjusts temperatures based on breakpoint analysis and meta data and also changes that occur without meta data documentation. I do not remember seeing the exact details or code used for the USHCN algorithm. If you have a link to it I would appreciate seeing it. I recall looking at adjusted USHCN station series (after USHCN applied the breakpoint algorithm) and still being able to detect breakpoints in difference series with adjacent stations.
I think your last sentence in your excerpted comment above indicates that breakpoints can be detected below the threshold level and are used if the breaks correspond to meta data.
I just re-read Menne’s 2009 paper and from it I think I can reconstruct what their algorithm is attempting to do. I am wondering if it can be simplified with fewer adjacent neighbors. More pairs would simply mean to me that you are going further away from the target station and looking at pairs with lower correlations.
Also in Menne’s video above I noticed that the largest errors were with the simulated series with numerous and small changes. With the false positive test certainly showing that the algorithm used was nopt over sensitive I was curious whether the sensitivy could be jacked up a notch or two. I was not able to determine under what conditions these small errors were not detected. It would appear that some were.
ftp://ftp.ncdc.noaa.gov/pub/data/ushcn/v2/monthly/menne-williams2009.pdf
Kenneth, That’s the most constructive discussion yet. Despite the hooplah by Watts et al regarding the false positives, they are miniscule and the formula should be improved with sensativity in mind… sensativity is its weakness… but then it was designed to find the big fish (discontinuities)
Here’s an example of what one hopes the breakpoint algorithm will catch:
http://www.worldclimatereport.com/archive/previous_issues/vol7/v7n1/figures/feature_fig3.gif
These stations are the Washington DC National Airport, Woodstock, a small town in the Shenandoah Valley, and the station for Charlottesville at UVA’s Leander McCormick Observatory. Clearly most of the record indicates that DC experienced an urban warming relative to Charlottesville…until 1997 when Charlottesville’s temps shoot up to meet DC’s (something curious appears to have happened in the late 80’s, too, to make the temps very similar then.) which turns out to have the nice, neat explanation that UVA built a big new professional complex with a parking lot, just a half mile south of the Charlottesville station. Doubtless these sorts of shifts, with varying reasons, can infect the records of even seemingly pristine stations.