New Computer Fund

Sunday, March 1, 2015

Using Tropical Correlation to Isolate Cause and Effect

If you have followed some of my rambling you know that I have been focusing on the tropical oceans and specifically the Indo-Pacific Warm Pool because of the Oppo et al. 2009 reconstruction which appears to be very closely correlated with "global" climate.  Areas with high correlation can be considered good "teleconnections" when trying to reconstruct a climate history.  No one location has a "perfect" teleconnection so you need to look for all the exception to the rule before other people start poking holes in you theory.  So this busy chart is hopefully a step in that direction.

This compares what I call the Hot tropics, 10S-10N with expanding ocean areas plus a NH and SH 0-30 degree band.  Areas of de-correlation should be related to perturbations like Volcanic activity or changes in ocean circulation that could be related to sea ice extent or shifting of atmospheric circulation patterns.  Toggwieler, J. R. with the GFDL mentions the "shifting" westerlies or changes in the Inter-tropical convergence zones (ITCZ) as a larger than many expect climate variable.

As should be expected, the correlation decreases as distance from the reference zone increases.  What is interesting is that the Southern Hemisphere (30S-0), the darkest green curve, and the Northern Hemisphere (0-30N) light blue curve, maintain a good correlation with the hot tropics and surprisingly well with each other.

This is the same 30 year sliding correlation between those two regions.  The roughly 1913 volcano shows as a small spike and the major climate shifts (red) seem to be where they would be expected except for around 2005 there is an indication of some other shift.  That shift is in the warmer direction, but pretty short term so far.  The main reason I am doing this is to see how well the sliding correlation window aligns with perturbations and that one is a bit of a surprise.  The other black arrow appears to be volcanic related but Pinatubo in 1991 is missing.  Pinatubo may have had equal hemispheric impact so it wouldn't show up in this correlation.  Somewhat surprisingly, Pinatubo doesn't really stand out in the other correlations.  It may have been more of an atmospheric thing than an ocean thing.

In any case these correlations are interesting to me.  I had expected the shifts to stand out and there to be a long term improvement in correlation, but not the major shift following the 1998 Super El Nino.  However, that shift seems to be related to surface winds, shifting westerlies, agreeing with Toggwieler, who might be a good addition to a lot of folks reading lists.  What wind data there is though just indicates an increase from a low around 1930-1940 without much variation.

Land surface temperature though gets most of the press.  Since land and ocean should be closely coupled, there should be a fairly high correlation between land and ocean temperatures.  Using the same main ocean bands I substituted the BEST global land for the Sstv4 (10S-10N) original reference to get this group of correlations.

This is a bit more interesting.  The same 1913 volcanic perturbation results in the lowest correlation then the correlation increases with time.  That to me would indicate either long term persistent recovery from past cooling and/or issues with the temperature data sets.   BEST uses kriging though and with most older long term stations being coastal, long term persistence should have a greater potential, with the exception of recent higher latitude station made practical with modern equipment.

In any case, there is a perturbation, ~1913, apparently due to the largest volcano of the 20th century, that was not include in model forcing estimates that has an obvious impact on temperatures.

This chart compares the CMIP5 model mean ocean temperature (tos) for the 50S-50N with the Crowley and Unterman 2013 estimated volcanic forcing (scaled to tos) to show the obvious difference in estimated forcing/impact.  This uses just the Northern Hemisphere which had the largest volcanic forcing.

The average volcanic forcing is a better match for CMIP5, but the ~1913 difference is still obvious.

"Forcing" though is not as simple as it may seem.  "Atmospheric" forcing is the focus of the models and the huge heat capacity of the oceans can "make-up" for negative atmospheric forcing and buffer positive forcing.  Satellite and ARGO era measurements make up for the limitations of "surface" temperature now but that doesn't help in determining if there is long term persistence.

With actual temperatures you can get a rough estimate of the change in atmospheric forcing.  You can't attribute the change very easily, but you can estimate the magnitude of the change.  Since the Oceans and atmosphere de-correlate fairly often, this might provided some insight but is going to be controversial.  This chart is based on seasonal (not actual Tmax and Tmin used to estimate Tave) Tmax and Tmin values for the regions in the legend using hadcrut Ts which is in C degrees.  CRUTs isn't intended to be a "global" temperature product but it does try to capture maximum and minimum values for determining dates of first frost and such.  The change in energy associated with Tmax and Tmin is surprisingly consistent for these regions.  I excluded the highest latitudes because that is were CRUTs accuracy seems to suffers most as should be expected.  According to this, there wasn't much change in atmospheric forcing until 1985 which is a problem period due to changes in surface temperature instrumentation.  In any case, this seems to indicate that current atmospheric forcing is about 4 Wm-2 above the 1901 to 2013 average.

While some might not like the CRUTs, Berkeley has an actual temperature estimate with about the same results.

The Northern Hemisphere is where most of the land is of course.

The Southern Hemisphere indicates less land forcing since BEST includes the Antarctic which may not be all that useful due to its high altitude and ultra cold temperatures.  The advantage of CRUTs was being able to mask regions which is difficult to do with BEST considering an absolute temperature.  All these used the same 1900 to 2013 mean and seasonal Tmax and Tmin not the actual Tmax and Tmin used to estimate Tave.  Note that the legend on the last two chart have BEST gbl instead of NH and SH as indicated in the chart titles which are correct.

The actual Tmin is currently being questioned due to variation the the nocturnal boundary layer as subject for another day.

The ~1913 volcanic perturbation shows up most strongly in the BEST NH Tmin energy approximation and BEST SH seems to indicate a continuous long term persistent recovery starting prior to that perturbation.   Remember all these used a 1901 to 2013 baseline.

Using the SSTv4 50S-50N temperature which had the best combined correlation with land temperatures I can also estimate forcing using the Crowley and Unterman 2013 Volcanic combined with the BEST interpolation of CO2.  I used a 1.6 C "sensitivity" estimate for CO2 forcing as a reference to all forcing aligning with the peak values.  This assumes that the evident cooling was natural/volcanic.  The volcanic forcing does not line up well with temperature (in this case converted to energy) response.  Assuming all the natural/volcanic cooling has completely recovered, the current atmospheric forcing is about 2.5 Wm-2 greater than it was in circa 1880.

Using the same baseline for comparison with "known" forcing equaling about 2 Wm-2 and SST for the 50S-50N region starting below the 1901-2013 mean, about half of the forcing indicated using the Land temperature data could be recovery from a long term persistent cool period just as indicated in the Oppo et al. 2009 Indo-Pacific Warm Pool reconstruction.

I am positive that this rambling look at various estimates of changes in atmospheric forcing isn't going to sway any believers, but it might inspire some to redo this in a more easily followed manner with more precise citations and proper jargon.  From my perspective, not including the more recent volcanic forcing estimates and not focusing on ocean and regional response has lead many modelers down a garden path.  It might be time to review some of the previous work of others, like Herbert Lamb.

Surface Temperature Product Issues

I have discussed this in numerous posts but thought I might as well consolidate some of that here again.  All of the land "surface" temperature products produce a "global" result that is very consistent with each other using different data sets and methods.  That is a great validation of the "surface" temperature.   Locally though there are very obvious differences in results for small areas that are "unique" in some way.

  Iceland is one example.  This chart compares Berkeley Earth with the Cowtan and Way kriging methods using Climate Explorer masking for the data.  I included the monthly data in the background and a 27 month 3 stage cascade  filter in the fore ground.  Berkeley produces a "country" product while Cowtan and Way doesn't, so I used Climate Explorer to hopefully produce an apples to apples comparison.  There is a fairly obvious difference in the results in Iceland for the two methods.

On a "global" scale the difference is meaningless, but if you happen to like maintaining Iceland's temperature record as pristine as possible, you might be concerned.  Any type of interpolation will result in some smoothing of data and very strong or "unique" data points will tend to be smeared across a larger area that normal "average" data points.  Kriging is a more "unbiased" way of interpolation, but that doesn't mean it is a perfect form of interpolation.

Iceland has the energy versus temperature anomaly issue I harp on in spades.  If you were kriging gold deposits you would want to include some indication of yield along with existence for any sample point.  If Iceland were the "mother lode", no other point would ever compare to Iceland.  So if you interpolate/krige a location north of Iceland that has a huge rise in winter temperature from say -36 to -26 C degrees with a southern area that has a small decrease in temperature from +15 to +14 C degrees, you would think Iceland warmed significantly.

The energy per degree change for the northern location would be about 3.2 Wm-2 and the energy for the one degree change for the southern location would be 5.4 Wm-2 excluding any latent energy change.  Energy wise, the southern location should be given greater weight during the interpolation.  For this particular case temperature anomaly would result in a 4.5 C change while energy anomaly excluding latent would result in a 3 C change.  So you need to include absolute temperature (or energy) when determining what a "peer" for Iceland should be for interpolation.

Since you don't have accurate latent data, you will never reach a perfect interpolation method, but including "sensible" energy will at least improve "pair" or "peer" selection if you want to produce a reasonable "local" product.  Once different methods start reducing "local" differences, there would be some progress made toward a consistent treatment of the energy that temperature anomaly is supposed to represent.  Then possibly "local" and "micro-site" impacts might be discernible.

Friday, February 27, 2015

da Mann of Natural Variability

Michael Mann has a post on Real Climate about his recent paper, Atlantic and Pacific multidecadal oscillations and Northern Hemisphere temperatures, Byron A. Steinman1, Michael E. Mann, Sonya K. Miller.

"In numerous previous studies, these oscillations have been linked to everything from global warming, to drought in the Sahel region of Africa, to increased Atlantic hurricane activity. In our article, we show that the methods used in most if not all of these previous studies have been flawed. They fail to give the correct answer when applied to a situation (a climate model simulation) where the true answer is known."

This paragraph caught my eye, previous studies are flawed because they fail to give the correct answers when applied to a climate model simulation.  To me that is replacing reality with a climate model.  The models should emulate reality not the other way around.

I have been struggling trying to find ways to compare model "output" with reality which is harder than I thought because the main model output "surface" temperature, doesn't really exist.  The models have a tas output, Temperature Air Surface, which would be the air temperature at some distance above the physical surface, including the oceans, and there isn't a reliable marine surface air temperature (MAT) product that matches the land based (Tmax+Tmin)/2.  There are MAT products, but they include night MAT only which would be similar to land Tmin, but land Tmin is considered to be unreliable.  I can create a model output, 70% SST and 30% Land air temperature, but that opens the door to questions about adjusting for potential temperature and sea ice area.  The simplest solution I have found is just stick to Sea Surface Temperature (SST) which has the majority of energy anyway.

I am also not a big fan of temperature anomaly.  It has its uses, but this is an energy problem.  CO2 increase should cause a small energy imbalance, on the order of 1 Wm-2, at the Top of the Atmosphere (TOA) which should take some time to reduce to zero if in fact it does reduce to zero.  Using various models the TOA imbalance has been estimated to be about 0.6 +/- 0.4 Wm-2 compared to a surface imbalance of possibly 0.6 +/- 17 Wm-2 (Stephens et al. 2012).

To get a better feel for "surface" energy I combined the Berkeley Earth land temperature in absolute temperature and the NOAA ERSSTv4 data using the 70% ocean and 30% land ratio to create a "surface" temperature.  As I mentioned though, that isn't a "real" model output as best I can tell.  That diverted me back to the SST alone as the best model to observation comparison I can find.

The following are ERSSTv4 temperatures converted to estimated energy and CMIP5 tos also converted to energy for various ocean regions.  The data was masked on Climate Explorer which may have errors, but looks pretty reasonable.  I don't have titles on these charts, but the regions are indicated in the ledges.

The 45S to 45N region makes the models look good.  This includes the majority of the ocean surface area and energy so the models should be doing a great job.

60S to 60N though isn't quite as impressive.  This indicates the models over estimated ocean energy by about 3 Wm-2 in circa 1915.

Since Mann is trying to show AMO, PMO (his term) combine to create a NMO or Northern Hemisphere Multidecadal Oscillation, let's look at the northern hemisphere oceans.

This region starts off with a 3 Wm-2 under estimation and ends with about a 2 Wm-2 under estimation in the region he is calling the "known" in his paper.

"We propose and test an alternative method for identifying these oscillations, which makes use of the climate simulations used in the most recent IPCC report (the so-called “CMIP5” simulations). These simulations are used to estimate the component of temperature changes due to increasing greenhouse gas concentrations and other human impacts plus the effects of volcanic eruptions and observed changes in solar output."  From the realclimate post with my bold.

Technical details concerning development of a 1200 yr proxy index for global volcanism,   Crowley and Unterman 2013, appears to be the newest reconstruction of volcanic forcing.  Their estimates are considerably different than those used in the "known" CMIP5 model runs with the most notable addition of a large volcanic forcing circa 1913 which is not in the CMIP5 forcing estimates and would help explain why the models consistently miss the 1900 to 1940 observations.  

In the southern hemisphere the models over estimate SST by about the same magnitude they under estimate the northern hemisphere.  In case you were wondering,

the tropics (15S-15N) might provide the best reference for how the models miss volcanic response and recovery.  That 1910s forcing is global based on the SST data and completely ignored in Mann's limited Northern Hemisphere "anomaly" comparison against a questionable "known" simulation.  

Now let's see how many of the "real" scientist pick up on this.