I am going to check our the hemicube
Hemicube and fisheye are both for integration on a surface. It sounds more like you're after an approximate where your object is just an arbitrary point/cube (I assume by element you mean a whole object and not a sample point on an object), in which case you probably want to be using a GBuffer cubemap and sampling that.
Although it's about light, thermal radiation is mostly the same (it just reacts with participating media [air] and in a multipass for convergence it has a loss - which you may not care about). Most exposure related things can be found in the same fashion just working with different values (kelvin instead of lumen, etc) and different tweaks (static absorption is strongest at perpendicular angles that along the vector to ground and never zero, etc).
The old Hugo Elias article is one of the best sources:
The Witness blog covered a fair bit on radiation: http://the-witness.net/news/2010/09/hemicube-rendering-and-integration/ as well as different projections.
Advanced Global Illumination 2006 and Practical Global Illumination with Irradiance Caching are good, but unless you're nuts about the subject only worth checking out from a library.
If you get lost, ask and I can toss a bunch of C++ code I use for doing this stuff (determining hemicube coordinates, multiplier map construction, and summation).
Wide but still less than 180 degree views tend to work fine though, something like an 18% inaccuracy compared to a hemicube sampling IIRC which is usually okay for anything non-life/ops-critical.