Better post something before I get completely swamped with jules' pictures...
I spotted this paper in the "in press" page for GRL about the same time as our last was going through the system. It came out some time ago, and got little attention despite being "highlighted" by the editor:
I spotted this paper in the "in press" page for GRL about the same time as our last was going through the system. It came out some time ago, and got little attention despite being "highlighted" by the editor:
They found that sea surface temperature data alone were not enough to make accurate predictions, but that assimilating monthly average subsurface temperature and salinity data significantly improved the accuracy of forecasts.Well, I suppose that doesn't sound so exciting really. However, it is making a significant point. This is based on the work I mentioned last year, when the first author visited Japan for a workshop. They have done a bunch of experiments to show that assimilating surface data generates a physically unreasonable response in the ocean circulation, whereas assimilating subsurface data as well produces a much better result. Of course, Keenlyside et al only (who are copiously referenced in this paper) used only surface data, and were taken to task by it on RC (1, 2). Ouch. Of course, with a different model and slightly different technical details, it is not impossible that K's method is useful. But they aren't prepared to bet on it, and neither am I!
32 comments:
Yes, you have returned just in the nick of time by my reckoning James.
What was she thinking?
Birds, bikes and exotic foodstuffs, fine. The "Toxic waste part of town", not so much.
If you average 17 AR4 GCMs, is it reasonable to say that whatever El Nino and other ocean oscilations are included in the models the timing of oscilations will be at various different times and will cancel out.
Is it therefore appropriate to take the average of the AR4 output and then make adjustments for such oscilations as we have experienced before comparing to reality?
Is your answer any different for ENSO (couple of year period and significant effect so lots of data) compared with your answer for PDO/AO/NAO/AMO and other longer term oscilations?
Err yes, OK. I should admit I am sort of wondering about the response to
http://www.drroyspencer.com/2010/01/evidence-for-natural-climate-cycles-in-the-ipcc-climate-models-20th-century-temperature-reconstructions/
crandles --- On decadal to centennial scales, climate time series of all sorts are approximately Pink Noise with the notable exception of the past 40 years or so.
Anyone feel like explaining the implications of different colours of noise as used in physics? The wikipedia link doesn't seem to give me any clue. Sorry I am so clueless.
Colors of noise
James, I can't see the paper, but do I take it correctly from the abstract that there's not as yet enough Argo data to do it for real? If so how much do they say is needed?
Oh, they also do it for real - the method is the basis of the HC "decadal" predictions - but this paper is purely methodological, so they use an "identical twin" experiment to eliminate the complications of observational limitations.
Basically, they are just trying to show that their method is better than Keenlyside, I think.
Correction - the HC decadal system uses ocean reanalyses that depend on a range of obs. This enabled them to do a number of hindcasts under the same conditions, and evaluate their forecasting skill.
This study uses various different full fields of data and it's not clear how much worse things would be with realistic observational coverage.
crandles: The other issue with the Spencer analysis is that my understanding is that there is reason to believe that the 1940s "blip" (which I think drives a decent part of Spencer's fitting routine) is an anomaly caused by WWII ocean measurement issues.
However, other than that, I think it is an interesting problem to try and figure out the historical underlying "internal variability" of the system in order to better understand what portion of the system is actually forced variability. I'm not convinced I'd trust Spencer's blog-science for this analysis (the man gets confused by the use of C13 in carbon cycle attribution because he doesn't realize that ecosystems and fossil fuels have the same C13 signature), but I'd look forward to a more competent analysis in the peer reviewed literature on the subject (but my guess is that it would be a hard problem).
-Marcus
Thanks Markus.
I am certainly not convinced I'd trust Spencer's blog-science for this analysis. I don't think 74% sounds plausible but I could easily be wrong.
"The other issue with the Spencer analysis"
I think I see several potential issues:
1. Are models tuned to the current phase of longer ocean cycles because the recent past is where we have the data from? (or have they all had long ocean spin ups done meaning they could be in any phase of an oscillation?)
2. Would you therefore need to remove this element from the models before doing the averaging?
3. There will be many ways of doing the lags and optimal combination of the ocean oscillations. How is overfitting avoided?
4. Adding an extra ocean oscillation data series just adds to the opportunities to match the deviations that have occured even if it has little effect on global average temperature.
I would think there is probably several more including the 40s' blip caaused by other reasons driving the analysis that you mention.
David, I did get as far as that page before asking what it meant. Sorry I am still nowhere near feeling I understand but an attempt would be:
"it has equal power in bands that are proportionally wide"
So because
"On decadal to centennial scales, climate time series of all sorts are approximately Pink Noise"
you expect to see equal power to affect global average temperature from oscillations that are roughly proportionally wide bands. So 10 to 20 year period ocean oscillations would be expected to explain as much of a global temerature anomaly as oscillations in a band of 20 to 40 year periods or 40 to 80 year periods. You simply wouldn't expect there to be some special frequency that could explain more of a global temperature anomaly than this would imply.
So if you come across a large temperature swing like medieval warm period to little ice age then recovery, this must be a result of forcing(s) not an ocean oscillation.
Is that something like what you were trying to say?
crandles --- Yes, regarding equal power that is approximately what is found in geophysical time series of many osrts. For centennial to millennial the noise redends, but never reaches red noise (1/f^2). So it is prefectly reasonable to attribute the sequence of MWP->LIA->after to simply this reddish pink noise. On the other hand, for that particular transition, there is W.F. Ruddiman's early anthropocene hypothesis, as given in his popular "Plows, Plagues and Petroleum". For similar transitions earlier in the Holocene (as found in the GISP2 ice core proxy record), of which there are very many, it is difficult to credit human causation. So while Ruddiman certainly has a good hypothesis, it works equally well to just view such jitters as more pink noise.
Of course on scales of tens of millennia and longer there is orbital forcing.
David Benson: crandles --- On decadal to centennial scales, climate time series of all sorts are approximately Pink Noise with the notable exception of the past 40 years or so.
You have a reference for this?
The reason I'm asking is because, if I compute the spectra for the fluctuating part of the temperature spectrum, I do not recover pink noise, but rather a series of discrete frequencies associated with normal atmospheric-ocean oscillations:
Figure.
Anonymous: crandles: The other issue with the Spencer analysis is that my understanding is that there is reason to believe that the 1940s "blip" (which I think drives a decent part of Spencer's fitting routine) is an anomaly caused by WWII ocean measurement issues.
The blip is also seen in the land record, (and has about the same magnitude) so that doesn't seem a very likely explanation.
Carrick: I know about the land blip, but I'm not sure that the land and ocean blips are actually consistent: I hate the whole email-stealing thing, but the following is the best explanation I know about:
http://www.eastangliaemails.com/emails.php?eid=1016
So Wigley's theory was that most land blips are 50% to 100% larger than corresponding ocean blips, but not in this case. Add to that that the 1940's "blip" is the place where global mean temperature observations and models are least consistent, and the lack of theoretical understanding for a reason for southern hemisphere cooling post-blip, and the world-wide disruptions going on around that time, and I think there is reason to think that maybe there's an observational problem there...
On the red/pink noise: have you applied that technique to any climate model output? I'd be curious as to how similar the climate model spectra are to observational spectra.
-Marcus
Carrick --- You time series for your figure ae quite short. Try 1850-1960 and you'll see what I mean.
Here is the reference:
Links between the Annual, Milankovitch, and Continuum of Temperature Variability
Peter Huybers & William Curry
Marcus: So Wigley's theory was that most land blips are 50% to 100% larger than corresponding ocean blips, but not in this case.
That's a good point.
I had thought of this too, since the relatively equal magnitude does suggest that the SST measurement might be inflated relative to the land temperatures.
I would have commented on this, but it occurred to me you need to look at the mean latitude where the warming on land is relative to the sea, since there is a strong latitude effect on temperature trend.
Certainly this is easy enough to check (makes note).
On the red/pink noise: have you applied that technique to any climate model output? I'd be curious as to how similar the climate model spectra are to observational spectra.
I've looked at a few models, but not many.
What I saw was the model output more closely resembles pink noise. There is some spectral structure in it, but I don't know if that is from the AO coupling i the model or from the forcings they are assuming.
As I understand the issue, most models have a very coarse spatial grid (e.g., 250-km x 250-km resolution), and for most finite-element codes you need at least 10 points per wavelength of an oscillation to accurately capture the physics of that oscillation. For an oscillation of e.g., 1000-km motion, ideally you'd need a 100-km x 100-km spatial resolution (or write a pseudospectral code, which doesn't suffer from this issue at least).
There are some interesting implications if it is true that GCMs (as they are usually run in any case) are not capable of accurately "capturing" short-period climate variability, in particular, it undermines attempts to compare short-term temperature trends to model outputs in an effort to test the validity of the models.
David Benson: Carrick --- You time series for your figure ae quite short. Try 1850-1960 and you'll see what I mean.
50 years is plenty long enough to test for spectral structure over periods of 10 years of less. By averaging longer, all you are doing is risking smearing out the spectral amplitudes associated with oscillations, which would happen if the frequencies aren't strictly stable. (In which case you'd be misidentifying frequency drift and/or frequency modulation as 1/f noise.)
But that is a substantially different point than arguing that climate fluctuations looks like pink noise.
Also here is a longer proxy series (vertical scale is temperature fluctuation. Temperature reconstruction based on a Greenland ice-core data from 553-1973.
Many other proxy reconstructions, including Mann 2009 also show spectral periods in these longer reconstructions. There are also peer reviewed publications where these various spectral peaks are reported (D'Arrigo, 2006 is one of these).
As far as I can tell, the pink noise meme is driven by people trying to use AR-based methods for modeling serial correlation.
I had an error in my link. Here's the relevant part of the post with the link fixed:
[...]
But that is a substantially different point than arguing that climate fluctuations looks like pink noise.
Here is the GISTemp 150 year average.
There is spectral smearing, but 2,5, 8, 50 year amplitudes at least are still evident. The 20 year peak is there, but in the shoulder of the 50-year period This and other peaks can be identified using fluctuation-based processing.
Also here is a longer proxy series (vertical scale is temperature fluctuation. Temperature reconstruction based on a Greenland ice-core data from 553-1973.
[...]
"or write a pseudospectral code, which doesn't suffer from this issue at least"
Most(? certainly many) atmospheric models are spectral, if that helps you...
James Annan: Most(? certainly many) atmospheric models are spectral, if that helps you...
Thanks, it does.
If you see anything else I get wrong, I always appreciate the feedback.
What's your opinion on climate models and short period (less than 10 year) climate fluctuation? Do they get it "right" , if not 'can theny get it right" or do they just "choose not to"?
Just to clarify, we are talking about the same method, right?
The pseudospectral method is different than the ordinary Galerkin-based "spectral element" method.
In principle the models could do some of the natural oscillations, but not very well and only for a year or three after careful initialisation. You certainly can't set one running in 1900 and hope to get a big El Nino in 1998, for example. This is just chaos theory. The 1940 bump (compared to model outputs) is thought to be partly natural variability, maybe partly observational error, and maybe partly still unexplained.
If you keep on reinitialising, you can track the historical record pretty well, but this is really using the model as much to interpolate the obs as to make a dynamical prediction. This is what "re-analysis" is all about.
(Away at a meeting so haven't chased up every detail, especially Spencer's latest.)
Re: spectral, I don't know, you'd have to go digging in the literature. I just know they call it "spectral". I expect it varies across these models too.
Carrick --- Thanks for the links; interesting.
However, I suspect spurious, especially a peak around 55 years. If one does a power spectrum of GISP2 for the entire Holocene, that goes away. More fun, perhaps, is the proxy for El Nino in
Variability of El Niño/Southern Oscillation activity at millennial timescales during the Holocene epoch in which a 2000 year cycle is claimed. I'm highly skeptical.
The 22 year cycle you have in your links might well be due to a Pacific Ocean effect related to El Nino; do you see it in the paper I linked just above?
James Annan: In principle the models could do some of the natural oscillations, but not very well and only for a year or three after careful initialisation. You certainly can't set one running in 1900 and hope to get a big El Nino in 1998, for example. This is just chaos theory.
Well that's true (to extent the climate really is chaotic anyway), but even if it's chaotic you should be able to reproduce the same spectra peaks, even if not the same exact pattern of coupled oscillations.
Re: spectral, I don't know, you'd have to go digging in the literature. I just know they call it "spectral". I expect it varies across these models too.
Yeah, the spectral method is just "old school" finite element.
"Pseudospectral" is a relatively new method (within the last 10 years), given the challenges in implementing a full climate model I would have been awed if anybody had already gotten it working, let alone if most or many groups had working climate code using it already.
Where it has been implemented, it has advantages in terms of speed and solves the afore mentioned problem with needing high spatial resolution and fine time steps to resolve physical oscillations. Basically it transforms the problem into wavenumber space then does the forward propagation there.
Anyways, I think the question of whether you can expect climate models to be able to produce ENSOs and the like is a big issue. If they can't, that tells yo something about the spatial and time scales over which you'd expect them to be able to accurately predict future climate.
David Benson: However, I suspect spurious, especially a peak around 55 years
No, it's a real climate oscillation as are the other ones. They can all be identified with components of standard atmospheric-ocean oscillations such as the PDO.
one does a power spectrum of GISP2 for the entire Holocene, that goes away
That's a problem with ice core data. As you go further backwards in time, you get progressive temporal smearing (I think it's due to diffusion of the O18 between layers of ice).
It's actually pretty easy to see. As you increase the time span over which you are analyzing the spectral content, the short-period amplitudes get progressively more smoothed out.
Plus there's the other issue, which is the question of how stable the period of the various oscillations is over millennia scale averages.
Carrick --- I've analyzed Alley's temperature data from GISP2 for the entire Holocene and the resolution over that period is good enough to detect periodicites down to 27 years (but not less). I assure you that fft on that data will produce simply a power law spectrum with no discrete spikes whatsoever. Indeed, using the Lomb periodogram technique, which is extremely sensitive, there is no statistically significant (to put it mildly) discrete periods between 27 and 300 year periods. What there is, but almost surely not statistically significant, is a bit of extra power from 30 to 90 year periods. I assume that is a result of ocean oscillations of one sort of another.
As for the purported 55 year cycle, from whence is it supposed to come? The 3.75 year cycle is a known recurring Kelvin/Rayleigh wave in the North Pacific; well studied. Around 8--10 years there are other components of ENSO and of course right next door to the sunpsot cycle. The 22 year cycle might be PDO, it shows up in analyses of the northern and equatorial Pacific. It seems much too long to be an oceanic Rayleigh wave and 55 years is completely improbable.
However, pink noise processes will produce what appears to be periodic behavior for awhile and then that goes away. In the ENSDO proxy paper I previously linked the wavelet analysis shows that clearly. I fear there is a tendency for people to "see" periodic behavior where in fact nothing but noise with hisotry is actually present.
David, once again the problem with GISP2 and ice core data in general is loss of temporal resolution in the ice core data as you move further backwards in time. This is discussed in the IPCC reports, for example:
It should be noted that the temporal resolution of the borehole estimates decreases sharply back in time, making it perilous to compare the shape of the trend shown in Figure 2.19 with better-resolved trends determined from higher-resolution climate proxy data discussed below.
You can confirm this yourself using the method I described above.
The longer oscillations are could driven by either salinity changes ("saline pump") or long-period solar oscillations but I'm no an expert on that.
But they certainly aren't thermally driven. Really long period oscillations may be associated again with changes in solar forcings (e.g., from orbitals shifts and what not).
You can analyze the difference between pink noise and a series of pseudo-stable oscillations using either the multitaper method (which I'm no expert on either, but there are a number of researchers including D'Arrigo who use this method effectively) or via fluctuation-based procesisng methods
The hypothesis that these fluctuations are due to a pink-noise process can be rejected using either of these methods.
Carrick --- Thanks, but you copied out about boreholes, which are much less precise than ice cores. I stand by my previous statement regarding GISP2 for the Holocene, only the top 10% or so.
If the 55 year cycle is real, it is certainly not commented about or documented anywhere AFAIK. It doesn't seem to show up in GSIP2 or the ENSO proxy I linked above.
From 10Be studies, the best explanation for changes in solar intensity is randomness; solar physicists have tried deligently to find anything more than that beyond the sunspot cycle and then also the very long term brightening.
Orbital forcing explains behavior at millennial and longer scales, not what I am currently concerned about. That is the scales from multidecadal to millennial which certainly appear to be best understood as pink noise (except for the mysterious 55 year cycle).
Figure 8 in
http://bprc.osu.edu/Icecore/masson.pdf
illustrates the pink noise power spectrum. There does not appear to be an outstanding 55 year oscillation in Antarctic ice core proxies for the entire Holocene.
The primary mechanism for a lot of these ideas about oscillations in the 55-year range seems to be wishful thinking. While such periodicity can't be excluded, the scientists doing serious work trying to find it will tell you that it doesn't in any manner cancel GHG-driven warming.
Well, it seems that AMO is about right for such a quasi-periodic oscillation. There is a fair bit about it on a NOAA site.
Post a Comment