I guess I'm not following what you see as divergence. The differences in how the data sets are scaled along the y-axis is largely a function of the baseline period used for calculating the anomaly. The baseline is different for different data sets. It's sort of exasperating but that's how it is. If you correct for that as an offset, which is done in the attached plot, you'll see that land data (HadCrut), land/sea extrapolated data (GISS) and the satellite data from both Remote Sensing Systems (RSS) and Univ. of Alabama (UAH) line up pretty well. Remarkably well actually, given that the land data is basically sticking a thermometer out the window and the satellite data calculates temperature indirectly from wavelength intensity. I agree there is a tendency for the satellite data to show a greater amplitude of both decadal warming and cooling events. I think that is your glitch. Why? I don't know-it could just be a statistical fallout of how the temperature values are indirectly calculated from the wavelength values. In the linked graph regression lines are also plotted. All of the slopes converge on ballparkish 2°C/century temp increase. I'm a bit disappointed that confidence limits (r2) for the regressions aren't provided on the WFT site. But hey, they are clearly trying to do a good job. Its from the same website you linked, woodfortrees notes section. If you want to have some fun with it, look up El Nino/El Nina events and superimpose them on the high/low chatter in the data. Mixing...
www.woodfortrees.org site (they want a nod obviously)
View attachment 497