[jules' pics] Guide to Hawaiian wildlife #3

Cute monkeys are everywhere on Oahu. The tufted great apes are ubiquitous - the little malnourished ones leap and swim in the ocean, while the big fat ones lie beached on the shore. In the trees, however, other species can be found:

--
Posted By Blogger to jules' pics at 1/11/2013 10:59:00 AM

[jules' pics] Guide to Hawaiian wildlife #2

Not all the fluffy aminals in Hawaii are quite as cute as they first appear. "Me bite You fall me bite you fall mebiteyoufall" "Ha ha ha hee hee heeee" "Me Top Turtle Now"

--
Posted By Blogger to jules' pics at 1/10/2013 10:54:00 AM

[jules' pics] Guide to Hawaiian wildlife

The giraffe and zebra roam the great plains of Oahu.


--
Posted By Blogger to jules' pics at 1/09/2013 05:00:00 PM

[jules' pics] Honolulu!

--
Posted By Blogger to jules' pics at 1/09/2013 09:12:00 AM

[jules' pics] New Year's Day

New Year sunrise at Kamakura beach was followed by some running along the Tamagawa. First there was the kiddywink racing. The adults seemed to be in more pain, but the faster ones spend more time in the air. We went back to pick up the tandem at the beach and found New Year sunset in progress.

--
Posted By Blogger to jules' pics at 1/03/2013 04:34:00 PM

Tamagawa 10k

A Happy New Year to all of my reader.

This year, I had a cunning plan for my New Year resolutions, making one of them to be that I run 10km in less than 40 minutes. Having entered a race on the morning of the 1st Jan, I had hopes of ticking that one off early in the year.

It didn't work out quite like that, unfortunately. The course seemed to be about 300m too long, according to both jules and I (note my watch shut off while waiting at the start line so missed part of the race), also subsequently checked with a measurement on walkjogrun.net. Although it was flat, the running surface was loose gravel, with a bit of grass, which made it heavy going (and possibly explains the mis-measurement, if they used a wheel). Given that, our times were fairly respectable, but that doesn't help with the NY resolution!

[jules' pics] Daibutsu

New Year's Eve - a time for reflection. We weren't actually the only ones there, but Kamakura is quiet. Everyone saves up their visit for the first 3 days of the New Year.

--
Posted By Blogger to jules' pics at 12/31/2012 08:32:00 PM

[jules' pics] Boxing Day

It didn't start out that well... But after a tasty business lunch in a basement, discussing a paper revision with a colleague also attending the meeting, we went home the slow way.

The light was a bit wild at the newly refurbished Tokyo station. Swung by the Imperial Palace The pale brown stuff below is grass. It always does this here, because it is so dry in winter. Found some end-of-year leaves (ginkgo) And some 10th month cherry blossom And then ran to meet James (who had been peacefully reviewing papers - his main occupation since August - in a cafe a couple of miles away), only pausing occasionally to photograph the architecture.

--
Posted By Blogger to jules' pics at 12/28/2012 05:15:00 PM

Can the Last Glacial Maximum constrain climate sensitivity?

You've had the appetiser from me, now for the main course from jules...

For many years now, a lot of researchers (ourselves included) have been working under the assumption - or at least hope - that models which better simulate the past are likely be more reliable for the future.

This assumption was dealt a bit of a blow by Michel Crucifix's paper in 2006, when he showed that, among the 4 GCMs for which outputs were available, there was no relationship between their sensitivity to the negative forcings of the LGM state, and their response to increased CO2. This suggests that any attempt to constrain the models directly according to their LGM simulations seems doomed to failure. However, Michel only had 4 models available to him at that time, and it would be hard to show significant results (the r-squared of a correlation would have to be 0.95 or higher to be statistically significant). Anyway, research continued, since paleoclimate simulations remain the only real opportunity to test the models' ability to simulate the large changes in climate that arise from large changes in forcing.

At the workshop we attended in Hawai'i early this year, there was talk of trying to write a review - or more accurately a "preview" - paper about the potential for using past modelled climates to constrain uncertainty in future climate in the context of the upcoming new multi-model climate ensemble (PMIP3/CMIP5). It occurred to me that people hadn't looked much at the spatial pattern of past and future correlation in the PMIP2/CMIP3 ensemble. We did this for our little MIROC ensemble a few years ago, and had seen quite a strong latitudinal variation, but I had been put off from doing something similar in PMIP2 by the slight inconsistencies between the model versions used for the past and future in PMIP2/CMIP3. I originally planned to wait until we had results from PMIP3/CMIP5 which should be a much larger and more consistent ensemble. However, for a preview paper I thought we could give it a shot with the old models, just to see what it looked like. There was the added incentive that writing the preview based on those models meant that in the future, when PMIP3 is complete, we will be able to see how totally wrong we were.

So, my approach was to bin the outputs of the PMIP2 simulations for the Last Glacial Maximum onto a 10x10 degree grid, correlate these local temperature anomalies with the models' climate sensitivties and hope to see a nice big correlation in the tropics where the dominant forcing for the LGM is due to GHG changes. And that's what we got. Only we also got a weird correlation the other way round in the southern ocean! That we don't understand, but at least it explains why there is still no correlation on the global scale despite the larger ensemble (7 models). In fact, when we looked again more carefully at Michel's paper, we were reminded that he had also looked at the tropics. He had compared past to future tropical changes which isn't quite the same as our analysis (we don't have spatial maps of future temperature for all models), but although his results do look weakly positive, his ensemble was too small for the result to be significant.

I've not really been so interested in climate sensitivity since we settled the matter to our satisfaction several years ago, but the rest of the world has been slow to catch up. So, it is natural to use this relationship, together with James' new LGM temperature estimate, to generate a sensitivity analysis. There are two fairly standard methods that people have used to do this.

One approach is to use the linear regression (and its associated predictive uncertainty) to directly map the observationally-derived temperature estimate (for tropical temperature) to climate sensitivity. This is basically what Boé et al did for sea ice. An alternative approach is to weight the GCMs according to how well they match the data. Both methods have their strengths and weaknesses, but in practice it doesn't seem to matter that much in this case. Our analyses point towards a moderate climate sensitivity of about 2.5C with a 90% range of about 1-4C, though there are some significant caveats in these results, which we hope the paper makes clear enough.

At some point in the summer, one of the IPCC authors contacted us to ask if we had any new papers on climate sensitivity. Meanwhile, the "preview" paper seemed slow in progressing, as these things tend to be. So, although our analysis was originally intended just as a modest constribution to that larger piece of work, we wrote it up as a stand-alone paper, submitted it to GRL, and somewhat to our surprise it sailed through review with only minor revision.

It's just been published on line, and is open access (which means everyone who is connected to the interwebs can read and download it). Alas, this is not due to any policy about-turn from the AGU, it's just that we had budget to burn.

How cold was the Last Glacial Maximum?

Several years ago, when we first got involved in paleoclimate research and were thinking about using the Last Glacial Maximum to constrain models, jules asked a respected researcher for an estimate of how cold the LGM was, in terms of a global annual mean temperature anomaly relative to the modern pre-industrial climate. Their surprising response was, "I haven't a clue". Of course this wasn't quite true, as the person in question was surely confident that the LGM was at least a degree colder overall than the preindustrial state, but not as much as (say) 20C colder. But they weren't prepared to specify a global average value, because the available proxy data were few and far between, and could not be considered to give good global coverage. 
 
When we looked at the literature, we found lots of analyses which looked at proxy data on a regional basis (such as SST in the tropics, and various polar ice cores) but not much relating to a global average value. Some people (including Jim Hansen) have given rough estimates of about 5-6C, and we used similar arguments to give an estimate of 6C when we needed one for this paper. But it was a bit hand-wavy. As for modelling results, in 2007, the IPCC AR4 gave an estimate of 4-7C colder than pre-industrial, which was based more-or-less directly on GCM simulations from the PMIP2 project.

Around the same time, various people (including us here) started more formally combining models with proxy data, by running model ensembles with different parameter values, and came up with similar estimates (e.g. here and here). So, a consensus view of around 5-6C as a best estimate seemed well established.

Since then, several major compilations of proxy data have been published (most notably, MARGO  for the ocean, and Bartlein et al for land), and although the MARGO authors did not generate a global mean estimate, there were strong hints that their data set was a bit warmer than the previously prevailing interpretation of ocean proxy data, especially in the tropics. When Andreas Schmittner and colleagues published their Science paper last year, their climate sensitivity estimate made the headlines, but it was actually their LGM reconstruction that was more immediately eye-catching to us. They fitted a simple(ish) climate model to these most recent and comprehensive proxy syntheses, and came up with a global mean temperature anomaly of only 3C (with an uncertainty range of 1.7-3.7C), which is far milder than most previous estimates. Irrespective of the resulting sensitivity estimate (which we'll return to later), such a warm LGM would be hard to reconcile with GCM simulations. Therefore, we thought it would be worth considering the LGM temperature field as a stand-alone problem. One obvious weakness of Schmittner et al's paper is that due to computational limitations, they only used a fairly simple climate model, which didn't seem to fit the data all that well. Our main idea was to see if we could do a better job using the state of the art GCMs. That requires a rather different approach, as we can't run large ensembles of these models.

Our resulting paper is still under review at CP, and available here. It has had useful and positive reviews, and I'm now revising it, but I don't expect the overall answer (which is an LGM cooling of 4.0+-0.8C) to change.

Our main result is based on a multi-model pattern scaling approach, based on a method I'd seen in the numerical weather prediction literature. It's actually just a simple multiple linear regression. The idea is that each model (from the PMIP2 database) is assigned a scaling factor in order that their weighted sum optimally matches the data (and hopefully, interpolates meaningfully between these points). This gave substantially better  results than either of the other two methods (which I'll describe below for the interested reader), and also worked well for a number of validation tests. The resulting reconstruction fits the data rather better than the results that Schmittner et al reported. The resulting estimated temperature anomaly is a nice round 4.0+-0.8C, with this uncertainty representing a 95% confidence interval. So that is a bit colder than Schmittner et al estimated, but also substantially warmer than most previous estimates. It's towards the low end of the range of GCM results, though with significant overlap. Our formal uncertainty range doesn't take account of the possibility that someone might come along in a few years and decide that these proxy data are all misinterpreted or hugely biased. For example, some people argue that the MARGO ocean data are too warm, especially in the tropics. We don't have a particular position on that, but intend to report some more sensitivity analyses in the final paper. Our analysis is just based on the need to interpolate into data voids, and the fact that different models suggest somewhat different interpolations.

And here is what you have all been waiting for - pictures of our reconstruction of the LGM anomaly, which are probably worth more than the 1000 words of this blog post. First, the surface air temperature anomaly, and then the sea surface temperature anomaly. These match closely over the open ocean, but at high latitudes the layer of sea ice insulates the ocean and allows them to diverge.

The dots are the proxy data points. Overall, the main features are very much as expected. Some of the smaller details are probably just noise, like the apparent "hole" in the top plot in the Southern Ocean at around 70S 30W, and also the N Pacific. The method generates uncertainty estimates which are particularly large in these regions. The slight warming in the Arctic SST (under thick sea ice) is in accordance with the proxy reconstructions, that in the south is more speculative as there are no data there

What this means for climate sensitivity, is left for a future post (some readers may be ahead of me at this point)...

Appendix: For anyone who is interested, here's a quick summary of two other methods we investigated. The paper has a fuller description.

We first tested the simplest method we could think of: direct smoothing of the data, as is commonly performed for modern temperature data by the likes of GISS and NCDC. However, the paleoclimate data are much more sparse and noisy than modern data, and the results looked pretty messy. Worse, when we tried sampling "pseudoproxy data" from GCM results (i.e., taking data from them in the same locations as the real proxy data) and smoothing into fields to see how well we could recover the whole field of GCM output, the results from this process had a huge overall bias. The basic reason for this is that there are no proxy data available in the areas which were covered by huge ice sheets back then but which are now bare (particularly North America and Scandinavia), and it is these regions which exhibit the largest anomalies for the LGM. So although we could smooth the real proxy data and obtain an estimated global mean temperature anomaly of 3.2C, we also knew this could easily be biased by a degree (or more, or less) compared to the real temperature. Bias-correcting this (based on the results generated by the pseudo-proxy experiments) gives an estimate of 4.1+-0.9C cooling. An ugly feature of the smoothing is that the spatial pattern generated is completely unreasonable, not showing any of the extreme cooling that must have occurred (simply due to direct albedo and altitude effects) over the large ice sheets that were in existence then.

So, we quickly concluded that we needed to use a model of some sort to interpolate between the data points. As an alternative to running an ensemble of simple climate models, we decided to use the state of the art PMIP2 model results which were generated by essentially the same full complexity GCMs which also contributed to the IPCC AR4 (CMIP3 project). We obviously weren't able to re-run ensembles of these models with different parameter values, so instead, we just used a simple pattern-scaling approach to fit them to the data. We think that this is pretty much comparable to changing the sensitivity parameter of a simple model - opinions seem to differ on to what extent this is true, but anyway it's all we could reasonably do with these models.

The results were markedly better than we got from the smoothing, but still rather disappointing. One minor curiosity, which we had not anticipated but which is obvious with hindsight, is that this pattern-scaling approach tends to lead to an underestimate of the overall temperature change. This is essentially due to regression attenuation which interested readers can google for more info (or look at our manuscript). Anyway, after bias correction this gives an overall estimate of about 4.5C of cooling, but the uncertainty is actually a bit higher than for smoothing, because different model patterns interpolate into the data voids differently.

So these two methods generate results which are compatible with our main result, but which are somewhat inferior.