Sadly the "doesn't survive compression/re-encoding" problem is kind of a show-stopper. I think steganography still has some important usecases, like covertly sending messages, but I don't think it's practical for "watermarking" images that are expected to be widely distributed.
btw, I noticed that the Merkle tree package you're using doesn't use a distinguisher bit in its node hashes: https://github.com/cbergoon/merkletree/blob/master/merkle_tr...
This can be a security issue in some contexts, since it means you can construct a leaf and a pair of nodes that have the same hash.