It's all very well bitching that the coverage map is at least partially based on predictive models, but if a particular location attracts no punters to make calls at it, then they can't very well say what the call quality will be at that point and are forced to rely on prediction models.
Additionally, anyone relying on a "green" pixel to say they will have good quality coverage in their basement faraday cage is also rather foolish.
There are tools around for the operators to produce these maps based on subscriber experience, but there is high variability in a variety of factors that influence the perceived worth of such maps:
a) geolocation accuracy - can be reasonably good in inner city, but out in the boonies if you only have coverage from one cell the accuracy can be a bit pants
b) statistical certainty - you need a goodly number of different mobes to make a statistical statement about the quality of the coverage in an area
c) environmental variables - leaves on trees cause problems, weather causes problems, so time of year will influence the results
d) overall network load - it's conceivable that a horde of punters doing something in a small area will have a detrimental effect on their surrounding area
Yes, you can normalise for all of the above, but it adds a lot of smoothing that makes the results less meaningful.
Short of the operators being forced to generate very detailed time series data for what happens to their network in every area, these maps are always going to have to be taken with a generous pinch of salt.
I agree that being able to do a network comparison would be quite nice, though.