The US Government shutdown delayed the release of the December GHCN station data. This also, perhaps surprisingly, also delayed the the Met Office/CRU results. So just how independent are they one from each other? Here is a comparison of the annual results from Berkeley Earth, GIStemp, my own 3d-GHCN, 3D-H4 and HadCRUT4.6, all on the same baseline of 1961-1990.
You can see that they all have the same shape but that they then begin to diverge after 2004. Why? GHCN V3 has 7280 valid stations and CRU have 7688 the vast majority of which use the same data, hence the reason for the delay in also releasing CRUTEM . The ocean surface data are also interdependent between HADSST3 and ERSST3, with only marginal differences. So what causes these apparent changes in results and trends ?
The basic difference is simply the way the surface weighted average is made.
- HadCRUT only average temperatures over locations where there are measurements based on a 5×5 lat,lon grid. The area weighted average of each cell is cos(Lat). This method has remained constant since 1990. You can argue that this is the only impartial choice since it avoids any interpolation.
- GISSTemp results in the steepest temperature rise because it assumes that every station is representative of temperature change within a 1200km radius. So the addition of new ‘rapidly’ warming stations in regions like the the Arctic have a far larger effect over adjacent regions when forming the average. My gut feeling is this method, originated by James Hansen has a warming biased.
- Berkeley Earth uses a least squares fitting technique based on an assumption that Temperature is a smooth function of position over the earth’s surface. So they also extrapolate into areas without data. Cowtan & Way use a kriging technique on the raw HadCRUT4 data, essentially doing the same thing.
- I use spherical triangulation of all station and ocean data over the surface of the sphere to cover all the earth’s surface. Each triangle has one measurement at each vertex and all the earth’s surface is covered. Nick Stokes uses a similar technique for TempLS.
Here is a comparison between the monthly temperature anomalies of HadCrut4 and GHCN V3 when calculated exactly the same way using Spherical Triangulation. Only about 5% of stations are different, but there remains a small difference in data corrections (homogenisation). HadCRUT4.6 uses 7688 stations and GHCN has 7280.
The new GHCN V4 provisional data has collected 27315 stations of which 17372 have at least 10y of data between 1961-1990, so eventually V4 should double the number of stations, although not dramatically increasing the geographic coverage. We can expect a bit of extra warming though since Arctic latitudes have yet more data. Here is a preview.
With so many new stations mainly in the Northern regions global temperatures will apparently ‘warm again’ once the main groups adopt it.
The December temperature rises by another ~0.05C above that of GHCN V3.
In conclusions all global temperature indices agree with each other in general, but that is mainly because they all use the same core station and ocean data. The main differences are due to how they spatially average the available data. HadCRUT4.6 is the most conservative because it only averages over (Lat, Lon) cells where there is real data. All the others extrapolate into regions without data. I think Spherical triangulation is the most honest because it works on the surface of a sphere, weighting each measurement equally.