Your algorithm must examine all possible parameters that could let it distinguish road from grass.
In terms of visual detection, this would include the differences we ourselves detect. Including color, shape, presence of many vertical elements (as grass has)...
There are also things which the human eye cannot see, but instruments can: example, heat signature.
There could also be differences based on size, uniformity of color, distance perspective... However these do not seem as though they are reliable characteristics to let you distinguish between a road and grass.
Seems to me you want to try to bring out the most easily detected difference. Therefore I would command the camera to filter out green. (Or should it filter every color >except< green?). I don't know what image you'll end up with.
You could try to program your neural network to discover the distinguishing characteristics. You would need to program it to examine every possible parameter. This sounds like an enormous task even for an expert programmer.