If I was looking to do this in volume I would try to find a material that was opaque in the visible yet transparent in the IR and another material that was opaque in the IR and transparent in the visible and arrange matching photodiodes/transistors. When the objects were brought in they would block the photodiode from its inband light and you could read the values on the PD's to determine what was in front of it.
I say in volume as it may be expensive to set up a run initially for the "water" and "food" objects.