I've found that using hashes and sending duplicates of the data seems to weed out any transmission issues in a one way system.
Consider an N bit stream of data, with a probability mu of any given bit being flipped. Then the probability of your stream containing an error is gamma := 1-(1-mu)^N. Since you're sending the identical stream twice, there's a 1-(1-gamma)^2 chance that your overall transmission is unrecoverable. The hash will tell you that the transmission failed, but not how to correct it. Furthermore, there's a probability that your hash has a bit flipped somewhere, too...
An error correcting code makes a guarantee that if up to m bits are flipped, the original message can be recovered exactly. A Reed-Solomon code can correct up to m errors while adding just 2m bits to the message length; even with a pretty conservative upper bound on the number of expected errors, this should be way less than 2N+k.
You get a couple of these Ethernet to 100FX converters: http://antaira.com/products/media-converters/unmanaged-conve...
Then you hook them up with just one fibre strand (instead of the usual 2). You obviously can't use TCP over that since that requires a 2-way connection, but UDP works just fine. You'll probably want to wrap your data in some error correcting code, too.