My job for the Antarctic paper so far has been simple. Present an area weighted reconstruction of the surface stations as a simple sanity check of the serious math reconstruction Ryan and Nic have been doing and I’ve been reading along on by email. So if you’ve been visiting once in a while you know I’ve presented several area weighted reconstruction’s. The algorithm I prefer so far is one which infill’s each satellite gridcell by the closest station which has data for that year I’ve created several ‘dozen’ versions, the latest one worth publishing was shown Some Other Area Weighted Reconstrucions and others were presented at Area Weighted TPCA Check .
Since the latest reconstructions utilize the noisy automatic weather stations which weren’t employed until after 1982 there are a lot of new series starting in the second half of the data. This creates some problems. First, we need to recognize that the Manned stations are more consistent and of a better quality than AWS stations which can become buried in snow for extended periods of time. A second issue is also important, each anomaly is centered about it’s mean. Average = 0. So if you have two thermometers which measure exactly the same values always and one starts 10 years after the other. Both have a mean of all their data of zero so if any trend exists thermometer 2 which measures exactly the same temperature as thermometer 1 ends up having a slightly different anomaly. When the two are averaged this creates a sudden step in the data.
Currently all presented plots do not correct for this factor. Since trends are fairly low, it isn’t a terrible method to use and we can expect that whatever corrected versions I come up with we’ll have a similar result. I’ve had several ideas of how to fix this annoying effect, some more stupid than others and none of which were worth posting on.
One particularly misguided method I used was to start at 1957 – the beginning year of our reconstruction – and offset existing Jan 1957 temps by zero. Then each time a new series came in I took the mean of all the other stations and applied that as the offset for the new data. I got a trend of 0.12 C/Decade which is out of whack completely and matches Steig et al. We know it’s out of whack by checking the surface station versions of the same algorithm which typically come to 0.04 or 0.05C/Decade. What the algorithm did by taking mean was to ignore the fact that the peninsula and Ross ice shelf are heavily oversampled in relation to the rest of the continent.
Ryan O made a great point that the offset in this manner assumes that when a new station is introduced and the mean is used, the temperature record for the new station prior to its existence is assumed to be the same as the continent. It’s an obviously incorrect assumption. In this post, I tried another method for correction. Since any offset assumes the trend of the new station prior to its own existence the next best option would be to choose the next closest station with existing data to compute the offset!
Well the algorithm I used was designed to calculated the offset between the new station and next closest station by using an average of the closest data points. I took the matrix of all ground station data on the sat grid 63 manned and aws stations and ran them through a filtering algorithm. This matrix FYo (filtered Yo)was stored separately and used only for calculating offsets.
squarefilter=function(Yo=Yo,windowsize=3)
{
FYo=Yo #allocate memory
for(j in time(Yo))
{
FYo[time(Yo)==j,]= colMeans(window(Yo,start=j-windowsize,end=j+windowsize),na.rm=TRUE)
}
FYo
}
FYo=squarefilter(Yo,2)
Read the rest of this entry »