Java's interface for reading filesystem folder contents is not very performant (as you've discovered). JDK 7 fixes this with a completely new interface for this sort of thing, which should bring native level performance to these sorts of operations.
The core issue is that Java makes a native system call for every single file. On a low latency interface, this is not that big of a deal - but on a network with even moderate latency, it really adds up. If you profile your algorithm above, you'll find that the bulk of the time is spent in the pesky isDirectory() call - that's because you are incurring a round trip for every single call to isDirectory(). Most modern OSes can provide this sort of information when the list of files/folders was originally requested (as opposed to querying each individual file path for it's properties).
If you can't wait for JDK7, one strategy for addressing this latency is to go multi-threaded and use an ExecutorService with a maximum # of threads to perform your recursion. It's not great (you have to deal with locking of your output data structures), but it'll be a heck of a lot faster than doing this single threaded.
In all of your discussions about this sort of thing, I highly recommend that you compare against the best you could do using native code (or even a command line script that does roughly the same thing). Saying that it takes an hour to traverse a network structure doesn't really mean that much. Telling us that you can do it native in 7 second, but it takes an hour in Java will get people's attention.