Hi @markus,
You were right about the 1/r^2 dependency. The source of my confusion was that I have done something very similar for the sun already, which is basically a point source at infinity. Hence, in this case, I can ignore the distance, the incoming ray density is constant and all rays parallel.
In the case of a radiating surface I will indeed need the surface normal and the depth information for each pixel to compute the radiative heat transfer between individual elements. I am going to check our the hemicube algorithm proposed by @AcidFaucent for that.