Here I re-calculated the sky brightness atlas using the original satellite data so that I can compare this to more recent data taken in 2001. For google earth and other users, the center of the upper left pixel is 125W,50N and the center of the lower right pixel is 66W,24N. The resolution is 1/120 degrees in both the longitude and latitude direction (same as the original atlas).
The DMSP satellite data comes from here. This DMSP data is special because it includes some observations taken when the satellite's gain setting is reduced so that urban cores are not saturated.
The light pollution model used in the original light pollution atlas was derived by Roy Garstang in the following two articles:
Garstang, RH: Model for artificial night-sky illumination, Publications of the Astronomical Society of the Pacific, 98 (601): 364-375, Mar 1986
Garstang, RH: Night-sky brightness at observatories and sites, Publications of the Astronomical Society of the Pacific, 101 (637): 306-329, Mar 1989
I programmed the Garstang model myself using the parameters given in the following paper.
You can compare my figure here to the original map for North America. There's very good agreement between the two, but if you look closely you can see differences. I've looked over my code many times and I do not think I made an error, moreover, there are at least two other sources of discrepancy: 1) Cinzano does not say how far he integrates the model up into the atmosphere. I integrate till convergence (i.e. integrating higher has a negligible effect) which is 100km. If Cinzano stopped at a lower level then light pollution would decrease faster with distance from a light source. 2) Cinzano does not how far a particular source effects the sky brightness in his code (i.e. does a source effect all points within a 300km radius or a 400km radius?). (You could have a source effect all points in the domain, but with the resolution here that would take a very long time to compute.) I tried testing a variety of values until it looked like the results converged. When I compare to Cinzano, however, I find the best agreement is for a "not yet converged" radius. The differences are small but still noticeable. I used the "not yet converged" radius here. Computer's have gotten faster over the last decade so maybe that is why Cinzano used a smaller radius.
The differences between these figures are well within the uncertainties involving the assumptions of the model.