Posted by Jeff Id on March 31, 2009
I ve been watching Phineus and Ferb cartoons with my 3 year old lately. The villan makes one machine after another with -inator on the end it cracks me up. Actually this post explores the blending of trends across the Antarctic from RegEm processing of he surface station and satellite data. It’s the Bass-o-matic for temp data.
This post uses a modified mapping script courtesy of RomanM which shows the correlation of satellite gridded data relative to a single time series across the antarctic. There are three columns of graphs below comparing surface station data with three varieties of the satellite data. I left all 42 surface stations in just as a reference, only 38 were used in the reconstruction.
1 – The first column is the raw Comiso AVHRR data as presented by Steig09. The data exists from 1982 to 2007.
2 – The second column is from the Steig09 reconstruction data, this data is also post 1982.
3- The third column is also from the reconstruction data, this is a comparison of the surface station data to the pre-1982. The satellite data didn’t exist during this period so it is entirely RegEM.
Remember the shape of the 3 pcs as plotted across the antarctic by CA HERE and the fact that there are only the first two effective PC’s as shown by this plot below.
In these plots below, data with less than 5 values for correlation in the time period was not plotted.
First look at the Faraday/Vernadsky plot. Column 2 has a clear vertical separation corresponding to a strong PC3 componenet yet in the pre 1982 data of column 3 the pattern shows a PC2 shape (If you are missing my point see the orange graphs in the right column at the CA link above). I just thought I’d point this out as the reduction in PC’s has an effect.
Another thing I noticed is that the correlations are much more pronounced in the pre-1982 reconstruction data as compared to the real data in the left column Leningradskaja is a good example as well as vostok.
Now many of us assumed that the peninsula trends must be blended across the antarctic in order to make the final trend happen. As far as station correlation is concerned, that doesn’t seem to be the case. Correlation is a funny thing though. Here’s a plot of 1982 correlation vs distance.
You can see that correlations as high as 1 exist over thousands of kilometers for the pre 1980 data. It appears now that the reason for this is a result of the 2 pc’s remaining in the RegEM portion of the satellite data – Column 3 above. It’s actually pretty easy to see the effect when you consider the multiple plots of large red high correlation areas in column 3 above where two thirds of the antarctic is dark red.
In my opinion, this doesn’t look very good as far as localizing stations but it’s not as bad as I had expected. We have only two areas of the antarctic in column 3 for determining when a station correlates well with gridded satellite data, so it’s just not possible to localize and properly area weight stations. It’s not like the peninsula has zero effect on 4000 km away either, in nearly every instance above there is a powerful negative correlation across the antarctic.
This post was a lot of work again but it really answers questions for me about the blending process. I think the next step is to try higher order PC’s and see what happens to the trends.