This is generally done in one of two ways.
The first is to represent your object with multiple simple volumes. A tree might be a cylinder with a sphere at its top, for instance. This is not particularly accurate but may be "accurate enough." If you don't need to really worry about colliding with the branches you could even just use a cylinder or even just a bounding-box enclosing only the trunk. You can be a little "sloppy" here; you only need to be so accurate.
The other option is to actually do a collision against a mesh, usually a low-poly mesh like you mentioned. This may be taken from one of your lowest LODs (levels of detail) or it may be especially made for physics. I don't believe XNA supports this. If not, you'll need to upgrade to a real physics/collision engine; there are a number of them but I'm not sure which are the better options for C# users. Typically though you'll want to stick to the first option as it tends to have significantly better performance, though some games do actually have need for mesh-collision (shooters, for instance, but then only for things that bullets need to be able to hit).