I’ve been away from WUWT this weekend for recovery from a cold plus family time as we have visitors, so I’m just now getting back to regular posting. Recently on the web there has been a lot of activity and discussions around the issue of the dropping of climatic weather stations aka “the march of the thermometers” as Joe D’Aleo and I reported in this compendium report on issues with surface temperature records.
Most of the station dropout issue covered in that report is based on the hard work of E. M. Smith, aka “chiefio“, who has been aggressively working through the data bias issues that develop when thermometers have been dropped from the Global Historical Climate Network. My contribution to the study of the dropout issue was essentially zero, as I focused on contributing what I’ve been studying for the past three years, the USHCN. USHCN has had a few station dropout issues, mostly due to closure, but nothing compared to the magnitude of what has happened in the GHCN.
That said, the GHCN station dropout Smith has been working on is a significant event, going from an inventory of 7000 stations worldwide to about 1000 now, and with lopsided spatial coverage of the globe. According to Smith, there’s also been an affinity for retaining airport stations over other kinds of stations. His count shows 92% of GHCN stations in the USA are sited at airports, with about 41% worldwide.
The dropout issue has been known for quite some time. Here’s a video that WUWT contributor John Goetz made in March 2008 that shows the global station dropout issue over time. You might want to hit the pause button at time 1:06 to see what recent global inventory looks like.
The question that is being debated is how that dropout affects the outcome of absolutes, averages, and trends. Some say that while the data bias issues show up in absolutes and averaging, it doesn’t effect trends at all when anomaly methods are applied.
Over at Lucia’s Blackboard blog there have been a couple of posts on the issue that raise some questions on methods. I’d like to thank both Lucia Liljegren and Zeke Hausfather for exploring the issue in an “open source” way. All the methods and code used have been posted there at Lucia’s blog which enables a number of people to have a look at and replicate the issue independently. That’s good.
E.M Smith at “chiefio” has completed a very detailed response to the issues raised there and elsewhere. You can read his essay here.
His essay is lengthy, I recommend giving yourself more than a few minutes to take it all in.
Joe D’Aleo and I will have more to say on this issue also.
rbateman (20:09:56) :
I believe the temperature record has been left in a lesser state than actually exists.
They didn’t just drop stations, they left them with gaping holes. Holes that I find out are sometimes plugged with data forgotten or a process never checked independently.
Don’t just believe.
http://clearclimatecode.org/the-1990s-station-dropout-does-not-have-a-warming-effect/
http://rankexploits.com/musings/2010/a-simple-model-for-spatially-weighted-temp-analysis/
http://www.drroyspencer.com/2010/02/new-work-on-the-recent-warming-of-northern-hemispheric-land-areas/
Check it.
Re: Tim Clark (Mar 8 06:17),
“Uh, Nick, I think you need to rethink that statement. We’re not talking absolute station temp here. If you drop stations that are showing no trend or a cooling trend and leave mostly stations that have a warming trend (airports), you bias the trend upward, regardless of gridding.”
No, what I said was “It only matters whether they are rising relative to that local long term mean.” Which seems to me to be exactly the same as you are saying. Of course, if you select high or low trend stations, that will be reflected in the results. But the D’Aleo/Watts report is all about selecting stations from warm regions, not with a warming trend.
In fact, one of the ironies is that it is cold stations that have the greatest warming trend (“Blame it on Canada”). So eliminating high-lat stations should make things cooler.
Nick: pre- or post-homogenization for those higher lats? Citation/link please.
I have come to the conclusion that the basic problem in the status of climate calculations is that the people doing them live in a linear world. Unfortunately the physics behind climate is highly nonlinear.
All statements saying that “anomalies reflect trends and correct biases” depend on there being one mechanism affecting the average temperature curve in each location, and on that being linear in the relevant variables.
Example:
A varying heat source in a closed room. Thermometers next to the source, and at various places in the room will show different temperatures, but a time plot of the temperature will show similar, within errors, variations in tandem with the source variations. Hotter close to the source, cooler far away. An anomaly over the average of each thermometer may be defined and it will show the same trend revealing the variations of the source and the anomaly can be averaged.
Add a second gradually moving heat source in the room and the same thermometers. Each thermometer will show different trends depending on the location with respect to the moving source and to the varying immovable source and anomalies over the average for each thermometer will be every which way.
The planet is much more complicated than this simple example. There are enormous motions of sea currents, air currents and jet streams , non linear and chaotic. These motions take bulk air with temperatures characteristic of their points of origin and distribute and mix it planet wide, non linearly. We then sit in some thousand places and take temperatures while varying non linear sources with different time sequences go about.
Again I will refer to the february anomaly plot: http://nsidc.org/images/arcticseaicenews/20100303_Figure4.png
It is worth studying it. There is an positive anomaly of 14C degrees in the arctic and a negative of 4C in Russia and all the rest in between. What has happened? Huge masses of hot air moved to the arctic where there are few storm systems and moved huge masses of frigid air at -45C south into multiple churning storm systems that mixed this cooling potential so that it ended up as a -4C anomaly in Russia. The overall positive anomaly reported is an artifact of ignoring the non linearity of the heat distribution system in the atmosphere and taking an average of apples and oranges.
And as I stated badly yesterday night above, by this miss match calculations it is quite possible for the heat content of the earth to be absolutely stable in time and the anomaly be positive.
Nick Stokes (20:53:22) :
Your argument has room for exchanging the dropped stations for the retained?
The results will be the same?
rbateman (09:43:28) : I have no idea how GISS manages to justify this:
http://gallery.surfacestations.org/main.php?g2_view=core.DownloadItem&g2_itemId=57846&g2_serialNumber=2
when the raw data for Ashland, Oregon looks like this:
http://www.robertb.darkhorizons.org/TempGr/AshOre1.GIF
Well, GISS never justifies anything 😉
(Think about it… I’ll wait… 2 meanings… )
But that one IS pretty bizarre …
I took the opportunity to use it as a ‘test case’ for my dT/dt method. Mine looks pretty darned good in comparison. I know, I ought to make a graph out of it (especially now that I’ve actually made graphs in a posting so everyone knows I can do it now and have the hardware 😉
But that box is presently shut down, so it’s “reboot and all that” stuff… (I hope to be out of the ‘reboot to change what you do’ process Real Soon Now thanks to some donated resources… – thanks guys, you know who you are 😉 but for right now, I’m going to post Yet Another Table Of Numbers…
The “dT” column is the ‘accumulated change of temperature’ while “count” will always be 1 for a single station. The dT/yr ought to match the change from one year to the next (so it ought to be easy to double check against the data chart). Be advised, the dT/yr is calculated from past to present while the ‘accumulated dT’ is shown present to past (puts the ‘baseline start of time’ at now).
Sure looks to me like a whole lot of nothing happening, right in sync with the base data you posted. Looks to me like GIStemp “screwed the pooch” somehow… But don’t worry, “The Anomaly Will Fix It” 😉
Mostly I’ve passed in country codes, but individual station IDs can be used too. Maybe I need to change that “For Country Code” to “For Stations in” … Hopefully I got it right from squinting at the fuzzy number on the graph in your link ….
It would be very interesting to run this on more odd cases. So much to do…
Re: Pamela Gray (Mar 8 21:00),
Pamela, a link
Ashland’s temps are not doing anything for one very good reason. Given that it is Ashland we are talking about, they don’t feel like doing anything ;>) cough cough.
Re: Alex Heyworth (Mar 8 15:14),
Thanks:
“WRT your second post, I would be convinced there was no issue IF I thought what GISS does in calculating anomalies was as simple as your example. As it is, I am about 3/4 convinced, but would like to see what Chiefio comes up with.”
A few points: CCC have the ability to do this test with GISSTEMP and their
reimplemented version ( which has been regression tested ) I believe that
they have already done this. ( carrot can you fetch nick barnes?)
Lucia has implemented the hansen method from A) the text B) checking the code to make sure she implemented it correctly. Zeke has implemented
a similar method, benchmarked against GISSTEMP. Grant Foster also has done a similar approach. CheifIo also did a limited test with GISSTEMP.
the difference was in the 1/100s place.
Like you I think the best test of the GISS method is to run the giss code.
Failing that you have a variety of approaches to the problem:
1) write code from scratch that matches the paper and the code ( lucia)
2) Examine the math describe in the paper ( nick stokes at first)
3) Toy examples of parts of the problem ( see mine here and Lucias first toy)
4) Independent solution with refernce to the problem ( Zeke &Foster)
6) independent database and approach ( Spenser)
All these aproaches and the limited example run by the chief show NO significant bias.
let me explain how people could approach this problem from an even more
conceptual standpoint. I think it helps frame the issue. Look at UAH.
from 1979-2010. You will see a trend, lets say its about .13C/century.
if 1979 is zero then 2010 is .4C ( trend wise) if dropping thermometers
on the land ( 1/3 of the total sample) dropped the land temps you would
see a departure from UAH only iff the drop was huge. lets just say the sea was .4C and land was .4C
if dropping thermometers dropped the land to .2C the total average would
drop to 3.3C. you would see that. ( .11C per century) If the sea was .4C and the land was .4C
and dropping thermometers reduced the land to .3C ( 25%) you see .366C
or a trend of .12C century. Lets suppose that dropping thermometers
hit the record by 10%. thats a big error, right. Then the land would show
.36 and the sea would still show .4 the average? .38C
very simply. It takes really big errors in the land record to move the overall average ( please dont argue about the SST record , its not germane). Now this doesnt mean that the land record should NOT be examined for improvements. But dont expect to find much.
There are at MOST a few 1/10’s of a degree on the table ( from 1979 on)
Mickittrick and Michaels estimate perhaps .3C total from 1979 to present.
In light of UAH that seems a bit much to expect.
So when people tell me they
More interesting is the record prior to 1979. More interesting is the phenomema Spenser desscribed in moving from zero population to small rural. Think about the early record. If you are looking for UHI and the spenser phenomena is right ( UHI is most visiable in the very low population to slightly bigger) Then finding UHI in the post 1940 record
is going to be increasingly difficult– signal thresholding. Finding it in the early record? neat problem.
so Nick,
I’ll make it easy for you, something you want to talk about:
is the anomaly for the dropped stations that same as those retained?
I won’t bother to press you about the actual temps for both.
Keith W. (14:12:13) : adjusted based upon the temperatures from any other stations within 1200 kilometers. This means that the gridcell containing Boston, Massachusetts, is affected by the temperature of the gridcell containing Atlanta, Georgia. While there would be a mitigating factor if there were temperatures from more Northernly gridcells, the higher latitude sites are decreasing.
BINGO!
There are some added bits of finesse to consider, but that is the basic issue and, IMHO, why GIStemp diverges from a “pure anomaly” result and why all the Hypothetical Cow “proofs” prove nothing. The Anomaly is a smoke screen behind which is hidden all the “other stuff” that STATION BIAS can influence. And with the Grid/box anomaly process coming AFTER all those other things, it can not “fix it”. No matter how hypothetically proven…
So we have, oh, a station in the high mountains 1000 km north of Boston gets dropped. Now a missing month of data (or in some cases, a missing year…) gets “filled in” based on a Carolina Beach. But in the baseline, the cold mountain is used, and in the present the Carolina Beach. “No Problem” as The Reference Station Method will “fix it up”; but we know from the FOIA emails that Gavin, at least, recognizes that bias can leak through…
So lets say that there was a Very Cold Year in the baseline ( It IS planted in a “swoon”) and that mountain was cold by 4 degrees compared to the average (high cold places DO have more volatility…) but now the data are from Carolina at a nice water moderated beach. Not going to get a -4 there, be lucky if it’s a -2 even on a similarly cold day / weather pattern … (Note: I’m saying cold compared to the average offset that GIStemp has computed in the Reference Station Method)
So we take these two “Boston” temperatures and find a -4 delta vs a -2 delta and we say “Boston” has warmed by +2 C anomaly. It doesn’t take many of these to get the average of them to be +1/2 C and discover “global warming”.
That is, IMHO, “how it works”.
Because the in-fill and the UHI and the homogenizing all happen before the GRID / Box anomaly step: STATION BIAS can cause errors in the TEMPERATURE DATA (that we see as those bizarre GISS temp graphs folks keep posting, like Ashland above) that can then make it into the anomaly maps ( that are made in a step AFTER those biased data / temperature graphs like, the Ashland map, are made.)
So I’m sorry, but the baseline DOES matter and the STATION BIAS does matter, and the station dropouts DO matter. Because they all feed into this error forming process. Prior to the grid / box anomaly production.
The ‘bits of finesse’:
As it does the UHI ‘correction’ it starts close and looks ever further out for a Reference Station to use. As you get further away, the correction gets worse. So while just down the coast from Pisa might work well, on the slopes of the German Alps works much less well. As stations are dropped, what UHI correction is done will become ever more dodgy due to looking ever further away for a ‘reference’.
Station drops matter.
A station with ‘too few rural neighbors’ will not get a UHI correction at all. So as the rural stations are dropped, there are an ever greater number of urban stations that have their temperatures simply ‘passed through’ unchanged. (Then a later bit of code drops them).
Station drops matter.
A USHCN station (with all those lovely adjustments in it) is “unadjusted” in STEP0 by comparing it with the GHCN record for the same station. If there is NO such station, the USHCN data is passed through AS IS (to be further ‘adjusted’ in the later GIStemp steps that are now “putting back in” the adjustments that were never taken out… a ‘double dip’). As GHCN stations plunged to just 134 in the USA, more USHCN stations are passed through “AS IS” and double dip adjusted.
Station drops matter.
A station will start looking for “nearby” stations for “in-fill” data via the reference station method. It starts nearby and circles out to 1000 km looking until it has “enough”. As stations are dropped, it must look ever further to find a reference station. The further away a station is, the less it will correctly provide fill-in and the more likely you have a bogosity. (the more likely it is to be in a divergent micro-climate zone).
Station drops matter.
I could go on, but I won’t.
The point: The actual code is NOT a hypothetical model and it does do things that depend on the actual number of stations available to it. As that number drops, those processes produce ever more broken results. Some of these have nothing to do with the “peer reviewed” papers that purport to describe what is happening, but have a lot to do with programmer choices. (Such as that “just pass the USHCN through unchanged if the GHCN is missing”… I’m sure that was not part of the paper describing how to ‘unadjust’ USHCN vs GHCN… )
So I’m sorry, but Station drops matter and station bias matters.
E.M. I plotted the dt and dt/yr data from above and put in a linear curve fit for each.
See here:
http://wattsupwiththat.files.wordpress.com/2010/03/ashland_dt-dt-yr.png
Carrot eater and Steve Mosher, thanks for your responses. CE, I was aware of the basics of the RSM from reading (so far) the first half or so of Hansen 1997. If all subgrids only had two stations to worry about, the method would be uncontroversial. It is what happens as a result of reiterating the process dozens of times that is more of an issue to my mind.
Steve, I am not overly concerned that there is likely to be a huge error in the global record to date as a result of the reduction of station numbers. As a number of commenters have pointed out, any resulting error is likely to be a reduction of the actual global trend, not an increase.
What it is a big issue for is regional variations from the global trend. This is important from the POV of assessing the actual impacts of GW. We are a long way from having any convincing picture of what regional impacts will be.
@ur momisugly Anthony Watts:
One heck of a lot flatter than the GIStemp results!
And a heck of a lot closer to the real data, if I do say so myself.
Thanks, many many thanks. It’s always nice to have a tool tested and shown to produce good results. 😉
If you have any other stations, or groups of station, where you would like such data for comparison, just send me an email. It’s about a minute a report to run them. Oh, and the code actually produces a file now with “Comma Separated Values” for Year, dT, dT/year, and Count. Makes it a lot easier to suck it into a graphing program… I don’t include the monthly data in the csv file as they are mostly for my own R&D interests.
On my infinite “to do” list is to use it to look at regions where GISS have rosy patches on their Anomaly Map and see what is actually there 😉
E.M.Smith
Purely conjecture, but Ashland is probably getting the benefit of the reverse sign UHI adjustment from Medford, which is only 21 km away.
Seen this sort of thing before in GISS. Odd inversion of adjustments.
Many of us ex programmers share a common approach. We want to see what happens to the numbers when GISSTEMP is run.
Other approaches, mathematical arguments, back of the envelop theorizing dont cut it for us. We want to know what happens in GISSTEMP.
made sense to me to ask for this
A while back I made this request.
http://chiefio.wordpress.com/2010/01/27/temperatures-now-compared-to-maintained-ghcn/#comment-2969
Run GISSTEMP with two datasets.
The dataset with the drop out
A dataset containing ONLY those stations that still remain.
this proceedure should give us a solid idea of how dropping or adding stations matters. The only time it matters is when the number of stations
becomes very small.. in the begining of the record.
As you can see from the Comments above EM also had this idea
but could not make it work.
ccc was able to make this test work.
ccc did that test: ( ccc have Gisstemp running as well. )
For those of you who know python and would like to mess around with GISSTEMP they have a great project.
Does station drop matter to the most important question?
the most important question is not how an individual sub box may get
impacted ( there are 8000 of them)
If you look you will find subblocks where the drop out can have an
effect ( especially if the trends are different between stations)
the question is this: what happens to the global average. Does it change
significantly?
Significantly?
Here is a nice link:
http://clearclimatecode.org/the-1990s-station-dropout-does-not-have-a-warming-effect/
you will see what the “difference” is.
And remember this is for the land. The land is 1/3 of the total. If you want
to change the GLOBAL average you have to move the land numbers by a lot.
Anthony Watts (23:08:26) :
Hi ANthony. Nick stokes has implement a process that will allow you to see how a station is adjusted and he output a list of stations for me.
have a look.
Paul Daniel Ash (18:54:01) :
yes there is a disconnect in how two different communities talk about
“significant difference”
Jan Pompe (16:28:40) :
steven mosher (13:55:32) :
“What do you think Nick?”
I’m not really sure that I care much for the opinion of a mathematician
******************
that’s ok. You know Nick and I disagree about plenty. But he has a blog.
he writes code and shares it. I can see what he does mathematically.
I dont always agree with him but his view is one that I would consider.
You see since he shows his math his views on other things dont matter
to me.
Sarnac (07:01:02) :
Simple airport temperature anomaly test …
That test was actually performed to show the impact of contrails
wikipedia. We discussed this at CA sometime ago, probably on unthreaded.
The grounding of planes for three days in the United States after September 11, 2001 provided a rare opportunity for scientists to study the effects of contrails on climate forcing. Measurements showed that without contrails, the local diurnal temperature range (difference of day and night temperatures) was about 1 degree Celsius higher than immediately before;[6] however, it has also been suggested that this was due to unusually clear weather during the period.[7]
Condensation trails have been suspect of causing “regional-scale surface temperature” changes for some time.[8][9] Researcher David J. Travis, an atmospheric scientist at the University of Wisconsin-Whitewater, has published and spoken on the measurable impacts of contrails on climate change in the science journal Nature and at the American Meteorological Society 10th Annual conference in Portland, Oregon. The effect of the change in aircraft contrail formation on the 3 days after the 11th was observed in surface temperature change, measured across over 4,000 reporting stations in the continental United States[8]. Travis’ research documented an “anomalous increase in the average diurnal temperature change”.[8] The diurnal temperature change (DTR) is the difference in the day’s highs and lows at any weather reporting station.[10] Travis observed a 1.8 degree Celsius departure from the two adjacent three-day periods to the 11th-14th.[8]. This increase was the largest recorded in 30 years, more than “2 standard deviations away from the mean DTR”.[8]
[edit]
anna v (21:08:02) : I have come to the conclusion that the basic problem in the status of climate calculations is that the people doing them live in a linear world. Unfortunately the physics behind climate is highly nonlinear.
All statements saying that “anomalies reflect trends and correct biases” depend on there being one mechanism affecting the average temperature curve in each location, and on that being linear in the relevant variables.
Anna, you are within a hairs width of what I’ve figured out. I cast it with a specific, but you have the generalization down cold. Bravo.
And now I’ll be quite on that point …
E.M.Smith (23:33:18) :
anna v (21:08:02) :
Ah haa! Nice one!
E M Smith
Exactly the same linear afflictions affect those calculating histroic sea levels
Reading the latter part of Chapter 5 of AR4 is a lesson in how many ways they find to say ‘We have no idea what is happening so have just extrapolated the already thin data using a straight line.’
Tonyb
E.M.Smith (22:24:26) :
A station with ‘too few rural neighbors’ will not get a UHI correction at all. So as the rural stations are dropped, there are an ever greater number of urban stations that have their temperatures simply ‘passed through’ unchanged.
As I understand it, an urban station with no rural neighbours will not “pass though” unchanged” but will simply not be used. There are many instances of trunkation of urban stations also where the urban record is long but the rural record adjusting it is short. The urban record is truncated to the date where it can be corrected by the rural neighbour(s).
Alex Heyworth (22:38:10) :
On the regional trend. yes.
I came up with another way of demonstration that there is no REAL
problem here. basically through comparing UAH LAND with GISS LAND
It’s interesting but I leave it for another day