I wonder if this kind of detection would work even better with hyperspectral cameras that have a dozen or more color channels instead of just the usual three. You could teach an AI to recognise the optical spectrum of specific paints used on munitions that would be undetectable to the human eye.

Such devices can be assembled at relatively low cost by combining multiple greyscale CCD or CMOS sensors with something like a set of dichroic splitters. Old TV cameras used to work this way, with three separate sensors for light split into red-green-blue channels. It's possible to split it further into finer divisions.