For GISS this month, I'll show their polar projection, because as Olof suggests, I think this is the key to the variability.
There are three main features. A big very warm region around the Arctic, but a large cool region around the Antarctic. In the context of present warmth, that cool Antarctic is cold indeed, and probably does more to bring down the temperature than the more intense cold in a band from Europe to China. So the end result depends on how the various indices add up these competing effects, two in regions lacking good coverage. GISS interpolates to give good weighting to the poles, and this boosts the Arctic effect. Normally I would expect TempLS mesh to match that, but apparently this time, it didn't quite. Or maybe it gave more weight to the Antarctic. TempLS grid definitely underweights the poles, and so the Europe cold effect dominates.
On this basis, I think it is likely that NOAA and HADCRUT will show a reduction for January.
To remark on where this stands with El Nino - on past experience the next four months or so are likely to be the peak. So the January rise is rather small. On the other hand, December (and October) had risen quite a lot, so it may just be a broadening of he peak.
Below the fold, I'll show the usual comparison between the GISS flat map and TempLS grid.
Here is GISS:
And here is TempLS grid:
It does look like Feb, on your data tracking, is going to separate to the warm side.
ReplyDelete"On this basis, I think it is likely that NOAA and HADCRUT will show a reduction for January. "
ReplyDeleteAlso true for the JMA-Dataset: http://ds.data.jma.go.jp/tcc/tcc/products/gwp/temp/dec_wld.html, its also setting a new record for january but in compare to december it cools down by -0.14K
Hi all,
ReplyDeletespent the last 1.5 days figuring what's actually going on, both in GFS and NCEP reanalysis, and updating things according to what I've found.
Turns out my GrADS script carried a bug ever since I've started plotting my maps back in 2012. Despite some intense checking back then, I overlooked one pretty nasty issue which caused quite some problems, amongst them the higher variability in GFS compared to GISS. What happened was that the global averaging only covered the eastern hemisphere owing to GrADS' tricky handling of 0-360degree longitude grids. Everything else (including the poles were fine). That explains why Eurasia's large winter variability also caused my GFS and NCEP anomalies to nose-dive or explode much stronger than GISS. Eventually, January 2016 made the problem so severe that I had to go back to the drawing board, i.e. defragmenting every piece of my postproc code again ... and sure enough ... there it was, the global averaging was wrong. Ridiculous how such a nasty bug can go undiscovered for years.
Since the anomalies would be different from now on anyways (and initial weighting was intentionally tailored to correct the actual code bug), I took the chance to re-tune GFS and NCEP anomalies to GISS on the basis of the last 4.5 years. This is how all the products stack up against each other: NCEP-GFS-GISS comparison
What stands out is an apparent "jump" of GFS in 2015. Thing is, GFS was updated mid-January 2015 and showed a tendency for a much warmer Antarctic region since ... in much better agreement with GISS. I therefore assume that the GFS bias has changed slightly with the update indeed. Hence I decided to apply a higher GISS-adjustment from Jan 2015 onwards. The actual new bias correction until Dec 2014 is -0.07K, increasing to -0.13K thereafter. The resulting difference for Jan 2016 between GISS and GFS looks like this: Bias GISS minus GFS
In fact, literally all the discrepancies that showed up before have now disappeared ... much to my delight. Not only is the necessary adjustment much smaller now (let alone the fact that the variance or standard deviation, resp, is almost identical now), but also would GFS have predicted the GISS outcome for January pretty well. Plus, the trend of all three products is almost identical after the "split" GFS bias adjustment as well. Noteworthy the pattern difference over the poles, where GISS just smooths things out with little regard to land or ocean, while GFS seems to keep things more physically realistic: Bias GISS minus GFS Arctic
Continue with part II ...
Part II ...
ReplyDeleteInteresting that GISS also seems to "smooth out" the cold Antarctic circumpolar SST anomalies (which ERSSTv4 probably shows), whereas GFS has no issue reproducing it: Bias GISS minus GFS Antarctica
Lastly, my NCEP/NCAR reanalysis anomalies suffered from the same problem. In reality, it's actually doing not any worse than GFS. The bias is even smaller. I only need to correct by -0.01K now and things look pretty already. January 2016 now remarkably similar in all three products.
So all my search for potential explanations for the discrepancies was misguided to some degree. In hindsight, I really wonder why I haven't been more suspicious about the differences between GFS/NCEP and GISS over the years. I guess I simply underestimated how good the three products actually are. Well, never underestimate the power of models ;) My apologies for the confusion that this might have caused particularly over the last couple of weeks during which (thanks to El Nino) have become pretty exciting. My updated "bug-free" product will hopefully provide much more reliable guidance from now on, even before the current month has ended.
@Nick: I will provide you with the link to the NetCDF files as soon as I find the time to incorporate it in the script.
KarSteN,
DeleteThanks for that. The varied conventions on lat/Lon are a pain. Sometimes 0-360, sometimes -90->90, sometimes 90->-90 etc. I like your idea of mapping the discrepancies - I should try that too.
I should try your idea of a fairly short anomaly period extended by GISS. It is a bit like what I currently do using a twenty-year period for NCEP/NCAR, which I think is about the good data duration. I then use GISS to extend to a 1951-80 period. The justification is that you need a long period to smooth out weather variations, but if you work out a bias, that compares the same weather in different indices, so the need for long period is less.
I see that in your new numbers, NCEP/NCAR was basically unchanged from Dec->Jan, and GFS went down by about 0.09°C. That seems to be in the range of other measures.
Thanks for working out how to make the NetCDF files.