In the 2007 analysis of the GISS dataset, Detroit Lakes was used as a test case. (See prior posts on this station here). I’ve revisited it in the BEST data set, comparing it to the older USHCN data that I have on hand from a few years ago.
First, here is a simple plot of USHCN raw and BEST versions. The BEST version is neither an anomaly series (like CRU) nor a temperature series (like USHCN). It is described as “seasonally adjusted”. The mechanism for seasonal adjustment is not described in the covering article. I presume that it’s somewhere in the archived code. The overall mean temperature for USHCN raw and Berkeley are very close. The data availability matches in this case – same starting point and same gaps (at a quick look). So no infilling thus far.
The Berkeley series is not, however, the overall average plus an anomaly as one might have guessed. Here is a barplot comparing monthly means of the two versions. While the Berkeley version obviously has much less variation than the observations, it isn’t constant either (as it would be if it were overall average plus monthly anomaly). I can’t figure out so far where the Berkeley monthly normals come from.
I then tried the following. I subtracted the Berkeley monthly average from each Berkeley data point and added back the USHCN monthly average. This yielded the following:
Figure 4. USHCN raw versus Berkeley (renormalized for each month)
The Berkeley data seems to be virtually identical to USHCN raw data less monthly normals that are different from normals of USCHN raw data plus annual average. The implied monthly averages in the BEST normalized data are shown below. The range of difference is from -2.27 to 1.41 deg C.
My original examination of Detroit Lakes and other stations was directed at whether NASA GISS had software to detect changes – a point that had been then been raised in internet debates by Josh Halpern as a rebuttal to the nascent surface stations project. I used Detroit Lakes as one of a number of type cases to examine this, accidentally observing the Y2K discontinuity. One corollary was that GISS software did not, after all, have the capability of detecting the injected Y2K discontinuity.
It would be interesting to test the BEST algorithm against the dataset with the Y2K discontinuity to see if they can pick it up with their present methodology. At first blush, it looks as though USHCN data is used pretty much as is, other than the curious monthly normals.
[Update: it looks like this data is prior to homogenization.]